Changelog

What's new in Yaji

New providers, new ways to work as a team, and the fixes that make each release steadier. Every release is here, newest first, with what to do after you upgrade.

Latest release0.71.0· Nov 14, 2026

Download

Nov 14, 2026Latest

Yaji 0.71.0

New

You can pinch on a touchpad or a touch screen to zoom in on the area under the pointer, as in Chrome. This works in the main window, in chat windows and in browser tabs. Diagrams, canvases and the image viewer keep their own pinch zoom. While an agent works in a browser tab, that tab does not zoom.

Improved

A new hosted server can start on a virtual server that OpenBot rents from Leaseweb, when one is free at the purchase. Such a server runs all the time: it does not stop after the last use, and it has 25, 75 or 150 GB for your files on Starter, Standard or Pro. When none is free, the server starts in a boat sandbox, as before. The privacy notice names the two locations (Amsterdam, and near Washington, D.C. for a purchase in US dollars or Brazilian reais).
A message that you send while an Antigravity agent works now goes into the running turn, as it does for Claude, Codex and Grok. It does not wait in the queue until the turn ends. This needs Antigravity CLI 1.5.0 or later. While the agent waits for your approval or answer, the message stays in the queue. On a joined server, the server must also be on this version.
You can buy a Starter hosted server again. It costs €30, $36, 120 zł or R$180 a month, and it has 25 GB of storage. A Starter server that you bought before keeps its price.
On the Canvas, you can drag a note, free text, a shape or a step by its text. A click on a step's text puts the caret where you clicked. A click on a note, free text or a shape that is not selected only selects it; a second click or a double-click puts the caret where you clicked. While a text field is open for edits, a drag in it selects text. A click on the empty Canvas closes the open text field.

Removed

Remove MongoDB Atlas from the Apps catalog. MongoDB lets only the AI clients that it registered sign in, and OpenBot is not one of them. Before, Sign in opened a MongoDB page that said "This Authorization request is invalid", and the app never connected.

Fixed

When the remote desktop of a server did not start because of an unexpected error, the app showed only "Request failed." The app now says that the remote desktop did not start on the server, and the server writes the reason to its remote desktop log. The server must also be on the new version.
The payment page for a license or a server shows the "I'm purchasing as a business" checkbox from the start, and the checkbox opens the business name and the tax ID (for example, a Polish NIP). Before, the checkbox stayed hidden until you typed a business name, so you could not find the tax ID field. The page also no longer asks for a full billing address when the payment does not need it.
The agent browser no longer replaces the language header that a site sets on its own requests, so the Oracle Cloud console loads its lists after sign-in.
When many Pi agents started at the same time on a busy computer, a message could fail with "Pi request timed out: switch_session" before the agent started its turn, and you had to send it again. OpenBot now gives a new Pi process more time to start.
When a second or later account of a provider, such as a ChatGPT account, was no longer signed in, its row stayed "Connected" and the agent stopped with "ChatGPT did not return an authenticated account." The app now shows the account as signed out, says "This ChatGPT account is signed out. Sign in again.", and the account menu has "Sign in again". On a joined server, the server must also be on the new version.
When a reply was read aloud at a speed above 1×, the voice could stop for some seconds near the start of the reply and then continue. The app now prepares shorter parts at these speeds, so the next part is ready sooner. Readings at 1× and 0.75× do not change.
Say "The server did not accept that sign-in." when an MCP server refuses a sign-in at its last step. Before, Supabase's refusal showed "The server answered 404. Check the URL", but the address was correct.
When a chat with a local model does not fit in the model's context window, the chat now tells you to load the model with a larger context length in Ollama or LM Studio, or to choose a larger model. Before, the first message to an agent could fail with only "The compaction request cannot be reduced further without losing the latest exchange or checkpoint" or "Nothing to compact yet".
When Gemini (Antigravity) cannot open the settings file of an account, the account now says so and tells you what to do: remove the account and add it again, or, for your first Gemini account, quit the other apps that use Gemini and open OpenBot again. Before, the account said that it was signed out, and "Sign in again" could not repair it.
When your ChatGPT plan limit is used but your account has credits, the chat now shows "Using credits" with the time that the plan limit resets. Before, it showed "Usage limit reached" while Codex kept working on your credits. The model picker no longer marks the model as stopped in this case. Members of a joined server see the same, when the server runs this version.
A browser tab that you do not see and that no agent uses now slows down, as a background tab in Chrome does. Before, its videos and animations played all the time, and one page with many videos could use more than two CPU cores. After a restart, it started again by itself.
An agent on Grok could get "Tool not found" when it asked you to take over a browser tab for a sign-in or a CAPTCHA. No sign-in card showed, and the agent said that it was blocked. The agent now gets the correct name for these tools.
When an MCP server asks for something that OpenBot does not show, such as a sign-in on a web page, the notice now names the server and tells you what to do next. When Codex asks you to reconnect an app on ChatGPT, the notice names the app. Before, the notice said only that a local plugin asked for an unsupported security hand-off. The notice now also shows in your language.
In the OpenBot browser, a part of a page that comes from a different site, such as a Google file or video picker, could stay empty, because the browser refused its request to use its own cookies. The browser now permits this request, as Chrome does.
When a sign-in field is in a frame from another address of the same site, as the Apple sign-in in App Store Connect is, an agent now learns that the field is in a frame and asks you to sign in with browser takeover. Before, the agent got "Element is no longer attached to the document" and tried other ways to fill the saved Vault login first.
An agent at Workspace only said that auto-approve was on, although it asks before each action outside its workspace. The agent now reads its correct permission level, so it can tell you why it asks.
Remote desktop of a self-hosted server or a Docker server showed an "×" in the center of an empty white screen, so it looked like the connection failed. The screen now shows a normal arrow pointer.
When an agent tries to change or delete a memory that is already gone, it now learns that the memory no longer exists and reads its memories again. Before, it was told that the memory belonged to a different agent, so it could not correct the memory and reported a problem that did not exist.
The model picker shows the models of the new version after OpenBot updates a provider's program, such as Claude Code, while OpenBot is open. Before, when the first read of the model list failed after the update, the picker kept the old list until you started OpenBot again: for example, it showed Claude Fable 5.1 and not Claude Fable 5.2. OpenBot now reads the list again after some seconds.
While an agent works, the moving light on its activity line uses less CPU. Before, the app redrew the text of the line many times each second for the full turn, also during a long command, so the fan ran and the battery drained. The light looks the same.
An agent that repeated the same failed step, such as an OpenCode web search with no search text, went on until you stopped it and used many tokens. OpenBot now stops the turn after 5 identical failures, and tells you to send the message again or choose another model.
When a local model could not summarize a long chat for OpenCode, each later message in that chat failed with "Compaction summary did not match the required template". Now OpenBot starts a new OpenCode session with the chat history and sends the message again. When it cannot send the message again, the chat tells you to try again or to select a stronger model.
An OpenCode chat no longer stops working after a model refuses an image or a file that it cannot read, for example a custom provider model that accepts text only. OpenBot now starts a new session with the chat history and tells you to send the message again without the file, or to clear Images at Accepts for a custom provider model. Before, some endpoints showed only their own error text, and each new message failed again until you chose a different model.
An agent can now read the text of a page that it showed before in a chat, for example a daily digest, so it does not repeat the same items. Before, the agent got only the title of the page from its chat history. For an agent on a server, the server must also be on this version.
Show why a channel request waits. Before, a request stayed queued with no reason when its agent worked in another channel, or when the channel already ran 2 tasks. Now the channel says which channel the agent works in, or that the agent starts when a running task ends. This also shows on a joined server and in the web client when the server is on this version.
Stop a routine run from the Routines canvas. While the run of the routine in focus works, Stop takes the place of Run now in the top bar, and it also shows in the routine details panel. Stop works for a routine that runs in the agent's chat and for one that runs in its own thread. The run ends as Cancelled, and the agents after it in the routine do not start. Before, the canvas had no Stop, and Run now could start a second run. Members of a joined server get Stop when the server runs this version.
An agent can now finish quietly when a teammate answers a request that a scheduled, script or webhook routine run sent, and the routine has nothing to tell you. Before, OpenBot refused the agent's "nothing to report" call, and the agent wrote a message in its chat with a notification. A Test run still shows the answer. A joined server must also be on the new version.
In the Natural response style, an agent chat shows the answer a few seconds after the turn ends, even when the chat did not get the final copy of the answer. Before, the chat could keep the answer hidden until you sent another message (#5847).
Connect YNAB in Marketplace > Apps opens the YNAB sign-in page again. Before, it stopped with "Redirect URI is not an allowed MCP client". When a sign-in server does not accept OpenBot at all, the dialog now says so in plain words. On a joined server, the server must also be on this version.
"Delete history" in the channel settings now also deletes the copy of the channel messages that each member keeps in its own channel work. Before, the channel became empty and the members forgot it, but these copies stayed in OpenBot. The members, their agents and the channel setup do not change. On a joined server, the server must also be on this version.
Open the Claude sign-in page on Windows when you connect Claude or add a Claude account. Before, on some computers no browser page opened, and the sign-in waited for 10 minutes and then failed.
You can sign in to an MCP server whose sign-in address moves to the same site with www. in front, for example Lunch Flow. Before, the sign-in stopped with "The OAuth endpoint redirected to …, so the credentials were not forwarded." A sign-in that moves to a different site still stops. On a joined server, the server must also be on this version.
In a voice conversation, the words that you say after a short pause are no longer lost. Before, a pause sent what you said, and the microphone stayed off until that part was sent, so the start of your next words, or all of them, did not reach the agent. Now the microphone keeps listening, and your next words go in the next message.
An agent no longer writes a note path from a notes app, such as Obsidian, as a file link. Such a link opened nothing and showed "Nothing exists at … in the agent workspace".
Show an image that you add to a message on a joined server as soon as its upload ends. Before, the message box waited for the app to download the same image back from the server, which took seconds for a large picture.
A Gemini agent whose earlier session the Gemini CLI can no longer read now starts a new session with its chat history and answers. Before, each message in that chat failed with "Failed to read session …; retry, or load it with session/load". An agent on a joined server gets the fix when the server is also on this version.
You can resize the right side panel when no pane is open in it: drag the left edge of the panel. Before, the panel showed only the list of panes, and its edge did not move.
The model picker shows a model that you add in GitHub Copilot's own settings, such as a model with your own key, within about 10 minutes. A model that you remove there also goes away. Before, while agents used GitHub Copilot, the list stayed the same until OpenBot or the server restarted.
In a thread, when a ChatGPT agent got a teammate's result and then ended its turn with no answer, the thread failed with "ChatGPT ended the turn with no answer". The agent now tries once more by itself, as it already did in its main chat. The thread shows the error only when the second try also gives no answer.
A ChatGPT (Codex) agent no longer offers to install a Codex plugin or to connect a ChatGPT app, such as Outlook Calendar, with a prompt that OpenBot cannot show. Before, the agent said that a connection prompt was in the chat, but the chat showed nothing. This is now also true when your own Codex settings turn these suggestions on. Apps that you already connected in ChatGPT keep working.
A video or audio file of up to 64 MB from a joined server plays in the desktop app chat, also after you move the slider. Before, the player could stop with an error or go back to the start, because the app sent the whole file when the player asked for a part of it.
When an agent cannot finish or stops a request from a WhatsApp contact or group on your personal number, OpenBot no longer sends "OpenBot could not finish this request." or "Stopped." to that chat in your name. The message still gets a reaction, and you see the failure in OpenBot. Your "Message yourself" chat and a number only for OpenBot still get the text. The server must also be on the new version.
In a voice conversation, a noise that continues, such as running water or dishes, no longer keeps the agent's spoken reply paused. After about two seconds with no words, the reply plays on. A sound that is transcribed with no letters or numbers is not sent as a message.
After you install a server on another computer with the install command, the server now shows in the server list in a few seconds. Before, the install card and the install dialog could stay at "Waiting for the command to run on your server…" until you restarted OpenBot.
Let you change the provider of an agent on a joined server while the agent works. The change waits for the next turn and the header shows "After this turn", as on this computer. Before, the server refused the change. The server must also be on this version.
A provider whose CLI starts slowly from a cold disk now starts. Before, the CLI had 10 s to show its version, and after 3 tries OpenBot stopped starting the provider until you refreshed the providers. Now, after one slow try, OpenBot gives that CLI up to 2 minutes. On a joined server, the server must also be on the new version.
When Google has no free capacity for a Gemini model, the chat now says that the model is busy and that your plan's quota is not the cause. Before, it said that a rate limit or the plan's quota was reached. On a joined server, the server must also be on the new version.
With an OpenRouter key, models with very long answers, such as Gemini 4 Carbon and Muse Wonder, now work with a balance of a few dollars. Before, OpenBot asked OpenRouter for up to about 944,000 answer tokens, OpenRouter refused the request, and the chat said that the account had no balance. A reply can now have up to 128,000 tokens. When the account still does not have enough credit for a reply, the message says so and tells you to add credit or select a different model. On a joined server, the server must also be on the new version.
An agent can save a skill from its own skill folder to the shared skill library. Before, the skill was saved but the agent reported that the installation failed, and each retry failed again. When an installation fails for another reason, the agent now gives that reason.
In a WhatsApp, Slack, Telegram or Discord conversation, an agent asked a teammate and got the result. When the agent then ended its turn with no answer, the result did not go back to the conversation. The agent now tries once more by itself, as it does in its own chat, and the conversation gets the answer. The conversation shows that the request failed only when the second try also gives no answer. On a server, the server must also be on this version.
When Kimi's server does not accept your saved Kimi sign-in, the chat now tells you to sign in with Kimi Code. When the server refuses a sign-in that has not expired, Providers also shows Kimi as signed out, with its sign-in button. Before, Kimi still showed as signed in, and each message ended with "Kimi ended the turn with no answer". A new chat without a saved sign-in also showed "RequestError: Authentication required" (#8555). On a joined server, the server must also be on this version.
Use a ClinePass subscription with Cline agents. The model menu now shows the ClinePass models, and they run on the subscription. Before, every Cline model used Cline Credits and failed when the credit balance was empty.
The automatic memory review now saves an agent's memories in the language that you write to the agent. Before, it often saved them in English, and then the agent started to answer you in English.
When OpenCode cannot read the reasoning of a model at a high effort, the chat now tells you to select a lower effort or another model. Before, the chat showed only "Internal error: reasoning start before end". This occurred with some Vercel AI Gateway models in turns with more than one tool step.
Compact context in a channel's settings compacts the members that can compact on request, such as Codex members, also after a restart or after some minutes with no answer. Before, it failed with "No member of this channel can compact its context now" in every channel. When no member can compact, the message now says why: no member has a conversation in the channel yet, or the members compact their context automatically. On a joined server, the server must also be on this version.
An agent that creates or changes a channel can now see the channel instructions when it lists the channels. Before, the list did not show them, so the agent told you that the instructions were not saved, but they were. When an agent sends a channel setting that the channel tools do not take, the error now tells the agent which settings it can use, so it can try again correctly.
On Windows, an OpenCode agent can attach files from its folder and the Shared folder to its answer. Before, the attachment often failed with "ENOENT from lstat", also when the file was there. If a path is still wrong, the error now tells the agent how to write it again.
A quiet routine run that has nothing to report no longer ends with "Grok stopped responding" after a 5-minute wait. When the agent sends nothing for 30 seconds after it reports no update, OpenBot ends the run as succeeded and shows no error.
A Gemini agent that waits for a long command now continues by itself when the command ends. Its next steps and their approval requests show in the chat. Before, the agent stopped with no message. About 5 minutes later, OpenBot stopped the Gemini process, and that cancelled the work. You had to send another message to get the result. On a joined server, the server must also have this version.
In a website editor, an agent's click on a link or button could go to the item next to it, and the click still showed as done. This happened when the page moved its text as the pointer came near, for example to show a block toolbar. Now OpenBot stops that click before the page gets it, finds the item again, and clicks it. On a joined server, the server must also have this version.
When a Kimi model in OpenCode could not read a PDF that the agent opened, each later message in that chat failed with "contains an invalid part type: file". Now OpenBot starts a new OpenCode session with the chat history and tells you to send the message again without the file, or to choose a model that can read files.
A ChatGPT agent's helpers can now use OpenBot actions, such as sending a message to another agent. The action runs as the agent that started the helper, and shows in that agent's chat. Before, it failed with "The sending OpenBot agent is unknown." On a joined server, this works when the server runs this version.
When another agent, a routine or an outside chat starts an agent's turn, and that agent cannot change a workflow, it now asks you to send the request as a new message. Before, it could ask you a question, try the change again after your answer, and fail again, so you had to write the same request again.
When an agent or you save a new password in the Vault for one subdomain, such as billing.example.com, OpenBot now adds a new login. The login of another subdomain with the same username, such as status.example.com, keeps its password. Before, OpenBot replaced the password of that login, and the old password was lost. On a joined server, the server must also be on this version.
When an agent saved a new skill whose name differed from a saved skill only in its non-Latin text, such as "da Vinci 投放工作手册" and "da Vinci 回复格式与语气", OpenBot refused it with "A local skill with this name already exists. Revise it instead." That text was wrong, and revising would replace the other skill. Now the message names the saved skill and says to change the Latin letters or digits in the new name.
Keep a site signed in on a server after you sign in to it on your own computer, also when the site keeps the sign-in in the page storage and not in a cookie. Before, OpenBot sent only the cookies, so some sites stayed signed out in the server's browser, and the request showed as done. The server must also be on this version.
An agent can now give another agent a Pi model whose name has an @, such as an OpenRouter preset, as its model or as a backup model. Before, OpenBot said that the model was not available, also when the model picker showed it.
When you ask an agent to delete a memory that you set yourself, the agent now tells you to delete it in its Memories settings. Before, the agent tried other memory tools and then said that the memory could not be deleted.
Voice model downloads, for replies read aloud and for dictation, now use the proxy settings of your computer, as the built-in browser does. Before, OpenBot ignored these settings. On a network that opens huggingface.co only through a proxy, the download showed "Could not download the voice model. Try again." each time, but your browser could get the same files.
Dictation on the desktop now uses a model that knows your language before you choose one in Settings > Voice. Before, it always used Parakeet, which knows only 25 European languages, so speech in Korean, Japanese, Chinese, Arabic and other languages came out as wrong words. When the app or the computer uses one of these languages, dictation now uses Whisper large-v3 turbo. The first dictation downloads it (about 1 GB), also when Parakeet is already on the computer. To keep Parakeet, select it in Settings > Voice. A model that you selected before stays as it is.
A ChatGPT (Codex) agent no longer shows a red notice that multi_agent is deprecated when the Codex settings file has that setting. The agent works as before.
An agent can click a video player button or another web page control that shows its icon through a slot of a web component. Before, the click failed with "Target is covered by svg".
A Codex agent can now start a subagent that gets the context of the run in a routine run, as it can in a chat. Before, the subagent did not start, and the routine run showed a "no rollout found" error. On a joined server, the server must also be on this version.
When OpenBot cannot read Bitwarden, agents still see your saved logins from the OpenBot Vault, 1Password and Proton Pass, and they tell you that Bitwarden could not be read. Before, one password manager that could not be read hid all saved logins from the agent, with no reason. The Bitwarden page now also shows the problem and what to do, next to "Connected". On a server, OpenBot unlocks Bitwarden again at the next sign-in.
In your own WhatsApp chat ("Message yourself"), each message from an agent no longer ends with one to three empty lines.
When you ask a channel's lead a question while it waits for the sub-tasks that it gave to other members, your question now starts as soon as a running task ends. Before, it waited until every queued sub-task had started, so the lead seemed to ignore you. While one person's message already runs in the channel, the other work starts in its usual order, so sub-tasks still go on when people keep writing. A routine's request and a message that an agent sends keep their place in the queue. A joined server must also be on this version.
When an agent continued its plan by itself after a new tool was added, it sometimes forgot your request and asked you to send it again. Now the agent gets your request again when it continues.
OpenCode no longer gets a new empty session each time OpenBot reads its model list. These sessions showed in the OpenCode app with only "model switched" lines. This needs OpenCode 2.0 or later. Sessions from before this fix stay; you can delete them in OpenCode.
When an agent's browser tab was not used for some time, OpenBot unloaded its page to free memory and loaded it again on the next use, but did not tell the agent. The address stayed the same, so the agent did not know that its scripts and page state were gone, and its browser tests failed with errors it could not explain. The agent now gets a note that the page was loaded again.
An agent that tests an MCP server, such as Gmail or Google Drive, no longer shows a Sign in card that cannot work. When the sign-in server of the service does not accept OpenBot, the agent tells you so and does not ask you to sign in again.
A ChatGPT agent that writes its last words as a short note and then ends its turn with an empty answer now shows that note as its answer. Before, the chat or thread showed no answer (#8837). A joined server needs this version too.
When an agent cannot stop work that it gave to a teammate, OpenBot now also cancels that agent's requests that still wait in the teammate's queue, so they do not start later. When the user or a different agent also sent a message to that work, the agent now gets the correct reason: only you can stop it, with Stop agent in the teammate's chat.
When we moved a hosted server to a new machine, files in folders outside the home folder, such as Obsidian vaults in /srv, did not move with it. Now a move finds these folders first and does not start until they are copied too.
A channel now routes messages when its lead uses a Pi model from a Pi extension, such as claude-code/claude-opus-5-5. Before, each message failed with "Routing failed. Choose a member or try again." Channel summaries, memory reviews and profile texts with such a model work again too.
A ChatGPT (Codex) agent in a channel can post its task result again. Before, on a busy computer or server, the channel sometimes refused the result with "The channel assignment is no longer active.", and the task stayed on running with no result.
A tool of an MCP server that you signed in to no longer fails with "OpenBot could not reach the server" when one new connection to the server fails, for example on a VPN. OpenBot now tries to connect once more. It does not send a request again when the server could have received it.
A Grok agent answers again after its workspace moves, for example after a move to a new server with a different home folder. Before, each message failed with "Path not found.: No such file or directory (os error 2)". Now, when Grok does not have the saved session in the new folder, the agent starts a new Grok session and keeps its chat.
When an agent gives a wrong id for a saved login, OpenBot no longer says that the login has no password. The agent now gets the correct reason and uses the id that the login list shows.
A routine whose task says to stop when the work is done now pauses itself when the work is done. Before, each run said that it could not change the routine and sent the same report again, until you paused the routine yourself. The routine stays in your list, so you can turn it on again. A Test run does not pause the routine. On a joined server, the server must also be on this version.
Scripts, MCP servers and provider settings that read a Vault API key from the shared key folder work again. Before, version 0.70.0 deleted the files in that folder when OpenBot started, and these tools stopped with no message. The folder again has each Vault key that is for all agents. A key that you give only to some agents is not in it.
A ChatGPT agent no longer compacts its context again and again without progress. Before, when each compaction left the chat too full, ChatGPT compacted after every step, and one chat compacted 559 times until OpenBot restarted. Now OpenBot stops the turn after the third such compaction and says what to change. On a joined server, the server must also be on this version.
Kimi agents can use an MCP app, such as Tavily, that you connected before OpenBot restarted. Before, the agent continued its old Kimi session, knew the names of the new tools and could not call them.
When Claude stopped a turn with "You've hit your monthly spend limit", an agent with backup models stopped and did not move to the next model. OpenBot now knows this message as a usage limit, so the agent continues on its next backup model.
An agent can now send a file or an image alone to a WhatsApp, Telegram, Slack or Discord chat, with no text. Before, the post failed when it had no text, and agents added a "." message to send the file. On a joined server, the server must also be on this version.
Agents can now use a Proton Pass login on every page of the saved site, as Proton Pass does. For example, a login saved for cloudflare.com now also signs in on dash.cloudflare.com.
A Slack workspace or a Telegram chat that sends its messages to a different OpenBot server no longer shows as connected. It now says that it is not linked to this server, and Reconnect links it again. This can happen after you copy the settings to a new server. A joined server must also be on this version.
A chat no longer stops working when a model behind OpenCode refuses it with "This request has more images than the limit of …". OpenBot now starts a new session with the chat history, as it does for other image limits, and tells you that the chat has too many images. Before, each new message failed again with the same error.
On a cloud server or a self-hosted server, each turn of a Claude agent with "Workspace only" failed with "Workspace only needs the Claude sandbox, which Claude cannot start on this computer" when the machine had no socat. The server now installs socat with its other packages. The server must also be on the new version.
The image viewer shows a small screenshot sharp when the display scale is above 100%, such as 125% or 150% on Windows. Before, the viewer made the image larger than its pixels and the text was blurry. The zoom level now counts screen pixels, so 100% shows the image as an image app does.
When an agent reads a video that has captions, such as a YouTube video, it now gets the full transcript from the captions. Before, agents got no transcript for videos longer than about 100 minutes, and only the first 30 minutes of shorter videos. When the agent asks only for the text, OpenBot no longer downloads the video. For a long video without captions, the agent is told that the video is too long, not that it has no sound.
A Claude agent now keeps the model in its profile. Before, an agent on the model that Claude Code showed as its default, such as Opus 5.5, changed to another model, such as Fable 5.1, when Claude Code changed its default after a usage limit. The profile still showed the first model. For an agent on a server, the server must also be on the new version.
Scripts and routines that read an API key for one agent at its old path, agent-api-keys/<agent>/<NAME>, work again after an update from 0.69 or earlier. Before, 0.70.0 moved these keys into the Vault and deleted the files, so these scripts stopped with a missing key. The path now links to the agent's copy of the Vault key, and comes back at the next start for agents that are already on 0.70.0. A server must also be on the new version.
The yearly price on the hosted server plans no longer breaks onto two lines in USD and BRL.
On a cloud server or a self-hosted server with Ubuntu 26.04, each command of a Claude agent with "Workspace only" failed with "apply-seccomp: ... caller must provide CAP_SYS_ADMIN". The server now adds one AppArmor rule for bwrap, and openbot uninstall removes it. The server must also be on the new version.
When you ask an agent to change the color of an agent's avatar, the agent now changes the color of the generated bot face. Before, it could replace the face with a one-color image.
A channel member that keeps working after a background command, such as a test run, now shows its progress and its answer under its task in the channel. Before, the task closed when the member's turn ended, and the work after the command showed only in the agent's own chat, so the member looked stuck. On a joined server, the server must also run this version.
Show when a Discord server moves to another OpenBot server. When you connect the same Discord server on a second OpenBot server, the first one stops answering there. Before, the first server still showed "Connected", and its agents got the error "unknown_guild". Now it shows "Needs attention" and tells you to select Reconnect to link the Discord server to it again. Agents get a plain message.

Security

Say in the Vault that every agent can read the API key files of the other agents, because all agents run as the same system user. The choice of agents for a key, and "Only this agent", control which agents OpenBot gives the key to. Before, the Vault said that agents read API keys from a private file, which was not true.

Nov 13, 2026

Yaji 0.70.0

Action needed after you upgrade

A server sends the full text of a long approval to the mobile app and the web client, so they can show it and allow it. Update each joined server to this version to use it there.

New

Restart a server that you host yourself from the app: as an owner or admin, select Restart server in the server menu. Before, only an OpenBot-hosted server and a server on this computer had this item, and a self-hosted server needed sudo openbot restart in a terminal. It works on the desktop and in the web client while the server is online. A server that does not answer still needs sudo openbot restart. The server must also be on the new version. In Docker, the container needs a restart policy, such as --restart unless-stopped.
In Server Settings > Vault, you can choose which agents can use each login, card, contact and API key. A new item is for all agents. An agent sees, fills and reads only the items that it can use. On a joined server this works in the desktop app and the web client when the server is also on this version.
Workflows have Image and Video steps. An Image step makes a picture with GPT Image (ChatGPT subscription) or Grok Imagine. A Video step makes a short clip with Grok Imagine. The card shows the file when it is ready. A video card has a Play button, and while the video plays, a bar with pause, the position, the time, a mute button and a full screen button. While the step runs, the card shows the time that passed. The aspect, length and resolution are on the right side of the card bar. The model menu shows only the models of a provider that is signed in, and it shows the subscription that pays for each file. Connect an Image step to an Image or Video step to use its picture: an Image step uses it as a reference. A Video step has an input for the first frame, the last frame and references; connect one picture to both frames to make a loop. Connect many pictures to References: Image steps, or Asset steps with pictures from the Library. To use a picture from your computer, click the upload button on an Asset step or drop the file on it. The picture goes to the server and stays in its Library. Use a PNG, JPEG or WebP picture. A Grok video can also use a voice. On a joined server, the server must also be on this version.
Change the agent defaults of a joined server in Server Settings › Agents, on the desktop app and in the web client. Owners and admins set local scripts, new chats, the plan card, "Steer agents while they work" and the commit co-author line, and which agents follow each default. The server must also be on the new version.
On a joined server, the desktop app and the web client can now set an agent to Ultrafast, and turn on fast mode for a Claude agent after the price confirmation. The server must also be on this version.
Owners and admins can set how many agent threads run at the same time on a server, in Server Settings › General › Threads at the same time: 1 to 128, or Unlimited. The default stays automatic: 2, 8 or 16 by the memory of a server, and no limit on a desktop computer. The server must also be on the new version.
Owners and admins can set how many threads each agent runs at the same time, 1 to 64, in Server Settings › General › Parallel threads per agent. The default stays 4. The server must also be on the new version.
A new setting in Settings > Agent dock, "Show agents with new messages", shows an agent or a channel that is not on the dock at the end of the dock while it has a new message or waits for you. It goes away after you read it. Right-click it and choose "Add to the dock" to keep it there. The setting is off by default.
The web sign-in opens in the language of your browser and has a language button to change it.
Connect Sunra from the Apps marketplace with browser sign-in, so agents can make images, videos, audio and 3D models with your Sunra credits.
Image and video steps on the Workflows canvas can use Sunra models: Nano Banana Pro, GPT Image 2, Seedream 5.0 Pro and FLUX 3 for images, and Seedance 2.0, Kling 4.0 and Veo 3.1 for videos. The models show after you sign in to Sunra in Apps, and the credits of your Sunra account pay for each file. A joined server must also be on this version to use them.
The model picker of image and video steps has a search field and a tab for each provider. When a provider has more than one account, each account has its own tab, and you choose the account that pays for the file. A step never uses another account. A joined server must also be on this version to show the accounts.
While the Canvas page is open, the page rail shows the Sunra credits. For a Sunra team member, a ring shows how much of the team quota is left. Select it to open Sunra in Apps. A joined server must also be on this version to show the credits.
Connect TryPost from the Apps marketplace with browser sign-in, so agents can list your social media channels, draft, schedule, queue and publish posts, and read analytics.
Image steps of GPT Image and Grok Imagine can make a picture with a transparent background. Choose Transparent in the new Background chip of the step. On a joined server, the server must also be on this version.
Paste a picture into the Canvas with ⌘V or Ctrl+V, such as a screenshot or an image that you copied in a browser. It becomes an asset at the pointer. PNG, JPEG and WebP pictures are supported.
The chat header shows the agent's title after its name, for example "Brock · Store manager". When the header is narrow, the title gets shorter first and then hides, so the name, the thread tabs and the model picker keep their room.
Add more than one Cursor account under Server settings › Providers › Cursor › Add account. Each account after the first uses a Cursor API key, and each agent can use a different account. Before, a second Cursor account replaced the first. On a joined server, the server must also be on this version.
Connect Adobe Analytics from the Apps marketplace with browser sign-in, so agents can find report suites, dimensions, metrics and segments, and run reports. Your Adobe administrator must give your account MCP access first.

Improved

An API key that you save "Only for this agent" from the chat is now kept in the encrypted Vault, for that agent. The keys that earlier versions kept in plain files move into the Vault when OpenBot starts. Agents still run as the same user of the computer, so this is not a strong wall between agents.
The key files that agents read have a new path. An agent gets the path each time it lists its keys, so a script that has the old path must use the new one.
An agent in a channel can now save to the Vault an API key that you wrote in your request in that channel. Before, it always refused and told you to add the key in Server settings > Vault. When the key is not in your request, or a key with that name is saved already, the agent asks for it in a secure form in the channel.
A request for an API key now shows the same card as a request for a login, card or address. Its "Add API key" button opens the Vault dialog, where you type the key and choose who can use it: all agents or only the agent that asked. After you save it, the card shows a green "Saved to Vault" badge. This card shows on this computer when the Vault is ready. A joined server shows the form as before.
A form that asks for a password or a secret shows a key icon while it waits.
The outcome badges of all chat cards now have the same size and fill.
In the full browser view, the message box is at the bottom centre of the page again, as wide as before. Before, it was at the bottom right (bottom left in right-to-left languages). The arrow in the message box still shrinks it into a small tab, which now stays at the bottom centre. Click the tab to show the box again.
Agent settings no longer show "Parallel threads". Server Settings › General sets it for all agents. An agent that already has its own limit keeps it.
The Add a server pricing page opens faster. The app loads its code before you open it. After the first opening, the page shows the last plans and prices at once and gets the new ones in the background. An order waits for the new answers.
The first screen of a new server welcomes its owner.
The agent settings, such as "Steer agents while they work", are now in Server Settings › Agents for each server, this computer too. Your settings no longer has an Agents page. Before, Your settings › Agents always showed the agents of this computer, also when a joined server was open.
Server Settings › Storage of this computer now shows the programs that agents left running, as it does for a joined server.
On the Settings and Server Settings pages, the close button is in the top right corner of the window. Before, it was farther from the edge.
The Canvas shows a picture with transparent pixels without a card behind it, at its own shape, so the canvas shows through. This applies to an asset and to the result of an image step.
The Library on the desktop app and on openbot.run/app shows a folder change at once when it comes from another device or from an agent. Before, it showed after the next storage scan. A joined server must also run this version.
The payment page for a license or a server now asks for a billing address and for agreement to the Terms of Service. It also has optional fields for your name, a business name and a tax ID.
When the main process freezes for 2 seconds or more in two minutes in a row, OpenBot now records which of its functions used the time. The summary stays in logs/stall-profiles.ndjson, and its five heaviest functions go out as one analytics event, at most once in 6 hours. It holds function names only, with no paths or data of the user.
The app opens as signed in from the saved session and confirms the session in the background. Before, the start waited for the account service to answer.
A joined server loads its workspace one time when the app connects. Before, the app loaded all the data two times, which sent about twice the requests to the server.

Fixed

OpenCode free models answer again when an OpenCode Go key is saved in OpenBot. Before, OpenBot also sent the saved key with the free models. When OpenCode refused that key, each free model failed with "OpenCode has no sign-in for the model provider "opencode"". A joined server gets this fix when the server is on the new version.
A new agent opens after you create it, on the desktop and in the web client. Before, the chat that was open before stayed open.
A sign-in with a saved or typed password in the agent browser no longer goes back to its first step on sites that answer on the same page, such as Tokopedia. Before, the page loaded again after the password, so an error went away, a sign-in that was still in progress stopped, and the agent was told that the password was sent.
Show a notice when a Codex model is at capacity. Codex then waits about 5 minutes between tries, and before, the agent showed as working with no output for 15 minutes or more. The notice shows at the first capacity retry that OpenBot gets, about 5 minutes after the send, and tells you to stop the turn and choose a different model to continue now. It shows on the desktop, in the web client and on mobile. For an agent on a joined server, the server must also be on the new version.
An agent that ends its turn in Portuguese with only an announcement, such as "Vou consultar o conector agora." or "Estou consultando o conector…", now continues and does the work, as it already did for English. English status lines such as "I'm checking the connector now." and "Let me query the connector." also continue now. Before, the turn ended with no result, and you had to send a new message. An answer that offers the work to you, waits for you or names a later time still stops.
M4A, WAV and AAC audio attachments now have a player in the chat, and "Open file" opens them in the default app. Before, they showed as plain files, and "Open file" only showed them in Finder or File Explorer.
On Windows and Linux, OpenBot now asks before it quits when you close its window while agents work. Select Keep Running to hide the window and let the agents finish; the tray icon then shows, and you open OpenBot again to show the window. Select Quit to stop them, or Cancel to keep the window. Before, closing the window quit OpenBot at once, and the agents stopped with no warning.
When an agent stops with "The agent could not continue. Try again.", the message now has a Copy details button. Before, the message did not show the cause, and you could find it only in a log file. Paste the copied text into a bug report to show what went wrong.
A failed reply between two channel members no longer leaves a "Cancel task" card after the request is answered. Before, the card stayed after a refresh, and an agent that you asked to cancel it said that the task had already ended. An agent can now also cancel such a card that an earlier version left.
Let devices connect to a server again when one device's connection stops during setup. Before, the server still showed "online", but no device could connect until the server restarted. The server must also be on this version.
When the shared database cannot start at first use, an agent now gets an error that tells it to try again, not an empty table list. Before, the agent could say that there were no tables, or make a second copy of a table that exists. Tables in agent settings now shows "Shared data is unavailable." in this case, not "No tables yet".
When Hermes cannot finish a turn, the chat no longer shows "Internal error: admitted_turn_failed". It now tells you to follow the reason in Hermes's reply, or to check the agent's MCP servers in Server settings › MCP, and then send your message again.
An agent can fill a sign-up form that asks for the password two times, from a saved login or from the password card. Before, the second fill emptied the first password box, so the form did not send.
Save on a skill card asks the agent of that chat to save the skill, also when you open another chat while it saves. Before, the request went to the chat that was open.
An OpenCode agent gets messages again after its folder moved. Before, each message to the agent failed with "session … does not belong to cwd". The agent now starts a new session with its chat history.
Open the workflow when you click its notification, so that you can approve or reject a step that waits for you. Before, the click only brought OpenBot to the front, and you had to find the workflow yourself.
When an agent lists the chats of a WhatsApp number, it now knows that the list has only the chats you allowed. When a chat is missing, it tells you to add the contact or the group in Marketplace, Apps, WhatsApp. Before, a personal number showed only "Message yourself", and the agent could not say why the other chats were missing.
A server counts file cache that it often reads as free memory. Before, a server with a small plan could show "This server is low on memory", refuse new browser tabs and hold scheduled runs, while the system showed enough available memory.
After an update of a provider such as Pi, OpenBot ran its own copy instead of the CLI that you installed. A launcher of yours, such as one that picks a Pi profile with its own models, then did not run, and those models were gone. The Runtime section of the provider now has "Use installed version" to run your own CLI again, and "Use OpenBot's version" to go back.
An agent can send a WhatsApp message to an allowed contact when it writes the number with a plus, spaces or dashes, or with the extra 9 of a Brazilian mobile number. Before, OpenBot refused the message and told you to add a contact that you had already added.
A channel routine no longer starts a new run while its earlier run still has open work. Before, the new run could take over the earlier run and stop its subtasks with "The parent request changed." The skipped time does not run later.
When another Codex app, such as the Codex desktop app, has a ChatGPT agent's chat open, a channel task that the agent cannot start now says that the chat is open in another Codex app and that you must close it there. Before, the channel said only "The provider could not start this assignment. Resume to try again.", and Resume failed again.
An agent that must not answer a WhatsApp contact, for example because its instructions say so, now stays silent. Before, the contact got the post "OpenBot finished without a written answer.". The message still gets a reaction, and the server must also be on the new version.
An agent can pay with a saved Vault card when the card fields are in a payment frame inside another payment frame, for example on some hotel and office booking sites. Before, the card fill always stopped with "Element is no longer attached to the document".
The live view of an agent's browser tab on a joined server stays open on a slow or distant connection. It shows fewer frames a second. Before, it closed after a few seconds with "The live view fell too far behind." The server must also be on this version.
On a joined server, Usage now shows the third usage bar, as it does on this computer: the OpenCode Go and Kimi weekly or monthly limit, and the Cursor plan total. Before, the server sent only two bars. The server must also be on this version.
An agent can fill a saved authenticator code from the Vault on more two-step sign-in pages, such as Okta ("Enter code"), Squarespace ("Authentication code") and Zerodha Kite ("External TOTP"). OpenBot now also reads the field's label and placeholder. It accepts a field named only "code" or "passcode" when the form names an authenticator app and does not name email or SMS. Before, these fields were refused with "the field targets are not code fields", and the user had to type the code.
GitHub Copilot agents can use Computer Use and the MCP servers that start on this computer. Before, Copilot did not load these servers. An agent with Computer Use on then said that it had no desktop controls and told the user to turn on Computer Use in Settings, which did not help.
OpenBot opens faster when your chats are long. Before, each start read every sent message two times, and checked all the past turns of each Claude Code agent again, also when nothing had changed. This made the window slow for some seconds after the start. Now a start loads the sent messages one time, and imports only the new Claude Code turns. A joined server gets this speed-up when it runs the new version.
When an agent drags an item in the browser, for example a module into a page editor, the page now gets the item. Before, some editors showed "Release to drop" for a short time and then added nothing.
An OpenCode agent's replies after a command that it ran in the background now show in the chat. Before, the chat stopped at the agent's progress note, and you saw the result only when you asked again.
A joined server whose connection drops again a few seconds after each connection no longer causes a new connection attempt every second. The desktop app now waits longer after each drop, up to one minute, until the connection stays up for 30 seconds. It also reads the server list again only after the first drop.
The web client no longer asks for a new connection to an offline server about every 15 seconds. It shows the server as offline and connects when the server comes back.
A server that cannot reach its Signal machine now waits longer between each request for a new one, up to 10 minutes. When the account service no longer accepts the server, the server stops these requests until it starts again.
A server sends a Live Activity update to a phone only once when the phone signed in again, and it stops sending updates after iOS ends the activity at 8 hours.
The web client checks a server that is not a hosted server for a hosted status only each 6 hours.
Sign-in to an MCP server that sends its answer as a form, such as Homey, now works. Before, the browser showed "This address answers GET only." and the sign-in in OpenBot did not finish.
When a service such as Omnisend does not accept OpenBot for its browser sign-in, the connect dialog now says that the service does not accept OpenBot yet and that your account is not the cause. Before, it showed only the text "unknown application: OpenBot".
In an agent chat, the agent's replies to a message that you sent while it worked now show below that message. Before, a reply that the agent wrote before its next step showed above your message, together with the other updates of the turn.
A new hosted server said that it was ready before it was online. It now says so when you can use it.
While a hosted server started, its notice showed an error code, Details and Retry, and you could not close it. It now shows only the start progress.
On a new server, a notice about an update restart showed while you set the server up. It now waits until the server has its first agent.
The provider list of a new server changed order while the providers loaded. It now keeps one order.
On a new server, a provider that needs a terminal sign-in on the host, such as Pi, showed as a main choice. It now shows under more providers.
On a new server, the free OpenCode models stayed selected after you connected ChatGPT or Claude, so the first agent used OpenCode. The provider that you connect is now selected.
On a server, the API key dialog of Gemini, Inception, DeepSeek and OpenRouter said that the key stays on this computer. It now says that the key stays on the host.
When you closed the setup of a new server, its update notice showed. It now waits until the server has its first agent.
"Open server" in the setup of a hosted server kept Settings or Billing open in front of the server. They now close.
A script in a skill stays executable after you install, revise, or turn on the skill. Before, the agent got "Permission denied" when it ran the script directly. Skills from GitHub keep their executable scripts too. A skill that is already installed gets the fix when you install or update it again. A skill that you saved before this version gets the fix when you save a new revision of it.
Pi shows as connected when a tool, such as a local model gateway, adds a provider and its key to Pi's models.json while OpenBot is open. Before, Pi stayed "Not connected" and asked you to sign in with /login until you restarted OpenBot.
The agent browser no longer reports a security check on a normal page of some sites that use Akamai, such as a flight list on ajet.com. Before, the agent waited, then asked you to take over a tab that had nothing to solve, and the report stayed after you selected Done. For an agent on a joined server, the server must also be on the new version.
An agent that reads a skill now gets each JSON example exactly as written, when the example has no secret in it. Before, a long number in an example, such as a 19-digit ID, came back with other last digits, and the spaces in the example were removed. An example that holds a secret, or a field such as token or apiKey, still has that value hidden.
"Reload page" in the browser now gets the page again from the site. Before, when a site sent a broken file one time, for example during an update of the site, the page could stay blank for hours, also after a reload. A reload by the agent does the same. For a browser on a joined server, the server must also be on the new version.
An agent can read and use a web page that hides all of its content from screen readers, or that makes it transparent, but still shows it. Before, the agent got an empty page with no text and no controls, for example on a site that left its page marked as hidden after a dialog closed.
When an agent reads a web page that has loaded but has no text and no controls, it now also gets a picture of the page. Before, it got only the empty page.
Sign in to the Microsoft 365 MCP app. Before, the Microsoft sign-in page showed the error "AADSTS900144", because OpenBot did not ask for the scopes that the server lists.
An agent that checks its HTML page in the light theme now sees the light chat background behind the page. Before, a page with a transparent background showed on a dark background, so the agent could not read its dark text.
A channel member that gave part of its task to another member and then posted its result in the same turn showed the same result again after each part was done. The member now posts its result once, after the other members return their results.
On a hosted server that runs on a virtual server, Storage now shows the correct free space of the plan. Before, it showed about 5 to 10 GB more free space than the plan had left. The server measures the files of its user, at most each 10 minutes. The server must also be on the new version.
The agents of a hosted server no longer get a wrong size for the system part of the disk.
Agents no longer see a provider that you did not set up, such as ChatGPT before you sign in, as ready to use. Before, an agent could choose its models for a new agent, and that agent did not answer. Now the agent sees that the provider is not connected, and for a provider that needs a sign-in it can tell you to fix it in Server settings > Providers. A joined server must also be on this version.
When ChatGPT (Codex) cannot start one of your MCP servers, OpenBot now names that server and sends you to the MCP settings. Before, some failures, such as "Unexpected content type", showed nothing in the app: the agent did not get the server's tools, and you could not see which server failed.
Codex agents compact the context of a chat that was idle for some minutes. Before, the compaction failed with "thread not found", and the chat continued with its full context.
An agent can now sign in with a saved authenticator code when the website has one code field and does not say how many digits the code has. Agents on a joined server get this fix when the server is on the new version.
Open a file or folder that an agent names in chat by its name or a relative path when it is in a folder that you added to the agent, or when the agent used the file earlier in the chat, also for an agent with Workspace only access. Before, the click showed "Nothing exists at … in the agent workspace". When more than one added folder has the path, the link does not select one of them.
When you post in a channel and the lead decides that your message needs no answer, the lead now answers it. Before, the request ended with no reply and nothing in the channel showed what happened. A routine's request or a message that an agent sends can still end with no answer. A joined server must also be on this version.
On macOS, Computer Use now becomes ready when you grant Screen Recording and Accessibility while OpenBot runs. Before, Settings > Computer Use and the agents kept saying that Computer Use was not ready until you quit and opened OpenBot again.
When a provider has no credits left and an agent moves to a backup model, the note in the chat now names the model that ran out of credits and tells you to add credits and select it again. Before, the note said only "Provider error", so the switch was easy to miss and its reason was not clear.
A long job that a ChatGPT (Codex) agent starts in the background on Linux or macOS now keeps running after the agent's command ends. Before, the job stopped at once and gave no result.
Hermes Agent connects when GitHub is slow or blocked on your network. Before, Hermes stayed in Error with "Hermes did not answer in time. The computer may be busy.", Refresh did not help, and the Providers page could show an older Hermes version. A joined server must also be on this version.
In Slack, Discord, Telegram and WhatsApp, a question about the status of work that a teammate did now goes to that teammate again. Before, OpenBot could give the teammate's earlier answer, for example "not published", after the work had continued in the app.
On a server that you made on this computer, agents can read, make, change and run workflows, and send feedback, when you ask in your own chat. Before, they refused your request as if it came from another team member.
On Linux, hiding, showing, minimizing or quitting OpenBot no longer shows "A JavaScript error occurred in the main process". Before, this error came each time after the desktop closed the hidden window that keeps agent browser pages drawn, mostly on Wayland.
Dropbox from the Apps marketplace now connects. Before, the card always said that the sign-in server does not accept OpenBot, and "Try again" did not help. To connect Dropbox to an agent on a joined server, that server must also be on this version. When the computer where you sign in has an older version, paste the address from the browser into the sign-in dialog.
An agent can fill a saved login again after it ran a script on a site that uses a service worker. Before, the fill was refused after each reload until OpenBot restarted. Now one reload by the agent is enough, and the agent is told which reload to use. For a browser on a joined server, the server must also be on the new version.
A browser profile on a server that uses a member's computer now says why an address on a local network does not open: the computer opens only public sites. Before, the page said that the computer does not answer, or showed an empty 502 page, and Browser settings marked the computer as not answering. The server must also be on this version.
An agent keeps a model that only its second account has, such as Claude Fable 5.2, when OpenBot starts again. Before, the agent moved to an older model of the first account, with no message. On a joined server, the server must also be on the new version.
A Hermes agent answers again after Hermes loses its chat session, for example after the Hermes gateway restarts. Before, each task from another agent and each message to it failed with "Internal error: not_found". OpenBot now opens the session again, or starts a new Hermes session with the chat history.
When a Claude Code Stop hook blocks a Claude agent's reply and the agent writes the reply again, the chat and the Slack, Discord and other channel posts now show only the new reply. Before, the blocked reply and the new reply came as one message, so the answer showed two times. The blocked reply now shows in the agent's thinking. An agent on a joined server gets this fix when the server is on this version.
Grok agents now get MCP servers that need up to 60 seconds to start, as long as the MCP connection test waits. Before, Grok stopped such a server after 30 seconds, so the test showed it as ready but the agent could not find its tools.
When Grok cannot start an MCP server, OpenBot now shows "MCP server not started" with the server name and the cause. Before, the server's tools were missing with no message.
ChatGPT (Codex) starts again on a computer with root certificates that Codex cannot read, such as the GOST certificates that Russian e-signature programs install on Windows. Before, since 0.68.0, Codex stopped at each start with "workspace routing discovery failed", and its agents did not answer. A joined server must also be on this version.
On Linux with Wayland, the OpenBot window now opens when OpenBot draws without the GPU, for example after an earlier start did not finish. Before, no window opened: KDE showed only a very small entry in the taskbar, and starting OpenBot again did not help. OpenBot now opens its window through XWayland in this case.
An agent finds earlier messages of its chats again when you ask it to look something up. Before, a search with several words found only messages that held those words together in the same order, and the list of chats did not show the chat that the agent was in. The agent then said that the conversation was lost. A joined server needs this version too.
After you change the model of a ChatGPT (Codex) agent, or after it uses a backup model, the chat no longer shows a red notice that the session was recorded with another model. The agent works as before.
Agents in a channel can read earlier messages again when one of them comes from a long agent turn. Before, the steps of that turn filled the history that the agent read, so the agent could not see the messages around it.
An agent no longer shows as working for hours after a quiet routine run, when the run's background subagents end after the run.
On the desktop and on openbot.run/app, a question card in a channel shows only while the agent still waits for the answer. Before, a card could stay at "Waiting for you" after the agent stopped waiting, and its answer failed with "This prompt is no longer active."
Chat history no longer saves unchanged messages again after the app reloads a chat in the background, so the database grows more slowly. An existing database keeps its current size.
When a Codex agent gave work to a subagent, and that subagent wanted to use a tool of an app or plugin, OpenBot refused the tool without asking you. The agent then reported that you rejected the tool. Now the request shows as a card in the agent's chat, and "Full access" allows it as usual. For an agent on a joined server, the server must also be on this version.
When a teammate ended its work on a request from another agent with no written answer, the agent that asked got a note with no details, also when the teammate ran tools or made files. It then sent the same request again. Now the note names the last tool steps, the files that the teammate made, and background work that still runs, and tells the agent to check that work first. On a joined server, the server must also be on the new version.
Show the full text of a long approval, such as a bug report for the OpenBot team, after the agent's status changes. Before, the desktop app and the web client could show only the first 240 characters and still offer Allow. A card that has only the start of a request now offers only Deny until the full request loads.
When an agent moves to a backup that has the same model on a different provider account, the note in the chat now names the account, for example "Provider error. Switched to Work · Gemini 3.8 Flash." Before, the note named only the model, so it read as if nothing changed. A joined server shows the new text after it updates.
Connect Railway in Apps opens the Railway sign-in again. Before, it stopped with "The sign-in server does not accept OpenBot as an app yet". For an app on a joined server, the server must also be on this version.
A GitHub Copilot agent now tells you to select a different model when its model does not accept the thinking setting. Before, the chat showed only "Execution failed: CAPIError: 400".
When an agent in a channel needs you to sign in to an MCP server, the channel now shows its Sign in card. Before, the agent said that a Sign in card was in the chat, but the channel did not show it.
Keep the iPhone Live Activity up to date while several agents run. Before, each change of state used the limit that iOS has for urgent updates, so a request or a reply could show late or not at all. Now only a new request, a new failure, a new reply and an agent that finishes its work are urgent. The server must also be on the new version.
A Workspace only agent no longer asks for approval for each file edit in a folder that you added to it. Before, a Cursor agent showed a "Change files" card for each edit there. A "Change files" card for one file now shows that file, not "Agent workspace".
On Linux, a start that was killed or lost power in its first 30 seconds no longer turns off the GPU until the next Chromium update. The next start tries the GPU again. Before, OpenBot then rendered on the CPU, which kept several cores and the fans busy.
Keep an agent's question open when the agent cannot wait for the answer, as with Cursor agents. Before, the question expired after about one minute, while you were still answering. Now the question stays until you answer, and your answer goes to the agent as a new message.
The Effort list of an OpenCode model no longer changes when OpenCode is busy. Before, a model that OpenCode did not answer for in time got the efforts of a different model, so the Effort box of an agent could be empty and its list could change each time the model picker opened. A model now keeps the efforts that OpenCode last gave for it. A joined server shows this after it updates.
When a page in the agent's browser stopped loading before it showed anything, the agent's next look at the page failed after 10 seconds. The agent now reads such a page at once.
When you ask an agent to send you a screenshot of a web page in its browser, the agent now attaches the image to its reply. Before, only the agent saw the picture, and the chat showed only text.
A paid OpenCode Zen model, such as Muse Ember 1.0, shows in the model list when you signed in to OpenCode Zen with opencode auth login, also when an OpenCode Go key is saved in OpenBot. Before, a saved Go key hid all paid Zen models.
A new OpenCode model shows in the model list a few seconds after OpenCode starts. Before, the list could miss a model that OpenCode released after its last start.
A message to a Grok agent no longer waits without end when the Grok process stops answering. OpenBot now ends that process after 5 minutes with no output, shows "Grok did not answer in time, so OpenBot ended its process.", and starts Grok again for the next message.
On the Linux AppImage, OpenBot keeps its files after it restarts itself (a restart when idle, or a restart for the dock above other apps). Before, remote desktop, Computer Use and the update check stopped working after such a restart, until you quit OpenBot and opened it again.
When a remote desktop component is missing, remote desktop reports that it did not start. Before, OpenBot showed a JavaScript error box.

Nov 12, 2026

Yaji 0.69.0

Action needed after you upgrade

The web client can now export the chat history, the channel messages and the server preferences of a server too, and it sends and gets a transfer file in parts, so the file can be larger than 100 MB in both directions. To export these parts from a joined server, update that server to this version too.
A hidden agent keeps working. It gets messages, runs its routines and does channel work while it is out of the list, so you can hide a backup or cleanup agent and leave it to its work. It shows in "Needs you" when it waits for an answer or an approval, and by default it notifies only then. Agents that you hid before this version had their routines stopped: OpenBot turns those routines off one time at the update. Turn on the routines of these agents that you want to run. Their queued messages run at the update. An import of a transfer file from an older version also adds the routines of a hidden agent turned off. It works on this computer, on a joined server and in the web client. A joined server must also be on the new version; an older server shows a hidden agent as paused, as before. An app on an older version still says that a hidden agent of a server on the new version is paused, but the agent works. To stop such an agent, turn off its routines.

New

When you add a custom provider and use Find models, the found models now open in a menu with a search field, as the add menu of the dock. Select all adds each model that the search shows, up to the limit of 64 models for one endpoint, and Deselect all removes them.
Teach a task: select "Teach a task" in the + menu of the composer, select "Start recording" above the page, then do the task in the agent's browser tab. OpenBot records the pages that you open and the controls that you click, not the text that you type. The stop button sends the agent one short message with the steps in a file; the agent saves the task as a skill and says in plain words what it learned. A recording with no steps sends nothing. It works on this computer, on a joined server and in the web client. A joined server must also be on the new version.
You can mute an agent from its menu in the sidebar. Right-click the agent and select Mute agent; Unmute agent gives it the notification level of the server again. The mute applies to you only, on each of your devices, and it is the same setting as Nothing in Settings > Notifications.
A routine, or an agent in your own chat, can read a connected Slack, Discord, Telegram or WhatsApp chat, for example for a summary each morning. WhatsApp and Telegram give only the messages that this computer received since OpenBot started. (#7793)
You can turn off the “Working on it…” message of a WhatsApp number. Open the WhatsApp page of the number and turn off “Working on it…” message: OpenBot then posts only the answer. The reaction still shows that OpenBot read your message, and a reply of stop still stops the request. For a number on a joined server, the server must also be on the new version. (#5711)
Pool the accounts of a provider: turn on "Pool accounts" in the accounts of a provider in Settings. Agents then use one account until it nears a usage limit, then the next account whose limit resets first. The pool keeps a part of each limit free, so a turn does not stop at a limit. Each account can have different limits, such as a five-hour and a weekly limit, or only a weekly limit. Select "Leave out of pool" in the menu of an account to keep it out. The model picker then shows one entry for the provider, with the account in use. The pool is off by default. It works on this computer, on a joined server and in the web client. A joined server must also be on the new version.
Connect hosts and devices in the Americas and in Asia-Pacific through new Signal and relay machines in San Jose and Sydney. A host now gets a near Signal machine, so remote chats there have less delay. Hosts move at their next reconnect. This needs no app update.
The Library now has folders. A side panel shows each agent with its folders, and a click on one shows only its files. You can make, rename, nest, move and delete folders, and put files into a folder by drag, from the "…" menu of a file or from the selection bar. A folder delete keeps its files and has an Undo. On a joined server, the server must also be on the new version.
The Library has three views: a mosaic of tiles (the default), a grid of tiles of one size with the file name below each tile, and a list.
An agent can put a file that it attaches into a Library folder. Tell the agent the folder, for example "save reports in Reports/2026", in a message or in its instructions. The agent makes a folder that does not exist. On a joined server, the server must also be on the new version.
Files that a deleted agent left in the Library are in a "Deleted agents" group in the side panel. You can open, download and delete them, but you cannot put them into a folder.
The Library is on the web client too, with its folders, views and picture viewer. A file that the browser cannot show downloads. The server must also be on the new version.
You can pin a message in an agent's chat or in one of its threads: right-click the message and select Pin. Agents can also pin a message, for example a decision or a checklist. The pin button in the chat header shows the pins of all chats of the agent, with the chat of each pin and who pinned it. One click goes to the message, also when it is old or in another thread: the thread opens directly at the message, and a button goes back to the latest message. A thread that is open in the side panel shows the message there. A chat keeps up to 50 pins. Pins also work in the web app. On a joined server, the server must also be on the new version.
You can set who gets each MCP server. The Marketplace page of a connected app or of a custom server has an Agents section, where you select the agents that can use it, and a Roles section, where Everyone and each role get No access, Can use or Can manage. A member with No access does not see the server. An owner or an admin sets this. On a joined server, the server must also be on the new version.
Disconnect the first Gemini account in Settings > Providers. Before, only the accounts after the first had an action that removed them. On a joined server, the server must also be on the new version.
A Grok agent can now make images and videos in the chat. The image or the video shows in its frame while Grok makes it, and then in place of the frame. A video plays in the chat, with its own controls to play, pause, move to a position, download and open it. On a joined server, the server must also be on this version. Before, the chat showed only the tool call.
Server settings > Transfer has a new "Copy to another server" action. Choose the server that has the agents on the left and the server to copy them to on the right, select the agents, and copy them in one step, with no file to save and open. The agents keep their workspace files, and their channels and the Shared folder are copied too. API keys, sign-ins, connectors, and chat history are not copied. It shows when you are an owner or admin of two or more servers. A joined server does not need a new version.
Add skills from GitHub. Paste a link to a public GitHub repository or folder into the Marketplace Skills search, and the skills in it show in place of the catalog. Add one to agents as you add a Marketplace skill. OpenBot keeps the commit that it read, and the skill page shows the repository and a link to it. OpenBot does not review these skills, so read a skill before you add it. On a joined server, the server must also be on the new version.
Server settings > Transfer can export a full copy of a server. Choose "Templates" for the agents and the server settings only, or "Everything" to also include the chat history, the channel messages, the workspace files and the server preferences. Each part also has its own switch, and the screen shows how much of the workspace files goes into the file. API keys, sign-ins and connector tokens are still never in the file.
A full copy now also moves: - the files of channel messages; - the run history of agent and channel routines, linked to the chats; - the shared skills that no agent uses (a skill with the same name on the new server is never replaced); - the server name and logo; - the model choices, browsing memory, Claude settings and machine alerts; - the notification levels and pins of each member who has joined the new server with the same account.
The owner's own notification level now moves also when the owner exports or imports from another computer.
The import has a new "Import server preferences" switch. It replaces the server name, logo and preferences of the new server, and each person's notification levels and pins, with the ones in the file. Only the owner can import the name, logo and the owner's own choices. "Import server settings" still only adds what is missing.

Improved

OpenBot uses Electron 45.3.0 with Chromium 156. Some sites refused the older browser version.
The message input is a pill when it has one line, and gets card corners when the draft grows. The queue, the waiting block, the notices and the @ list come out from behind it, as before.
The New chat button is no longer in the message input. New chat stays in the context ring.
The cards above the message input (waiting block, queue, notices, errors) use one row height, one text size and one text column, in line with the text in the input. A notice shows its title and its text on one line, with smaller buttons, so each notice with short text has the same height.
The cards above the message input open and close with one movement: a card comes out from behind the input and goes back down, and the other cards move with it. Before, most cards closed in one step.
The queue and the waiting block stay open while the @ list is open, so the chat does not move when you type @. The list opens over them.
The @ list closes with a short movement. Before, it went away in one step.
When a notice is the card next to the message input, the round ends of the input show the colour of the notice.
An error shows its reference code on the line of its text, so it has the height of the other notices.
The empty ring of the waiting block and of the open waits is easier to see.
A side thread, the side question panel and the Tasks page show their error and reply cards the same way: the card comes out from behind the input. The input in the dock chat and in the diagram chat is a pill too.
The input of the dock chat, the diagram chat and the side question panel is 40 px high. Before, it was 48 px in some panels.
The input of the dock chat, the diagram chat, the side question panel, a side thread and the Tasks page gets card corners when the draft has more than one line, and moves to its new height. Before, it changed height in one step.
The input of the Tasks page is one row: the agent, the text of the task, then Start. With attached files it has card corners, as the message input has.
Show the open server as a logo at the bottom of the page rail, below Settings. Click it to open the server menu on each page, also on Workflows, Tasks, Usage and Settings. Before, the server name was in the top row and showed only above the sidebar.
Show the New button at the end of the search field in the sidebar. Before, it was in the top row.
Make the top row the same on each page: Back, Forward and the sidebar button at the start. The chat tabs and the Tasks title start at the end of the sidebar. With the sidebar closed, and on Workflows and Usage, the tabs or the page title start directly after the buttons. The titles of Tasks, Usage, Schedule, Workflows and Settings have the same size, the same start and the same line as the buttons. A closed sidebar no longer makes the list of Settings or Marketplace narrower.
Add a sidebar button to the top row. It closes the agent list and opens it again, on the Agents page and on the Tasks page. The Tasks page no longer has its own button for the agent list and no longer hides the list each time it opens: it shows the list when the sidebar is open.
A message that you send to an agent while it works now steers the running turn by default. Before, it waited in the queue until the turn ended. To get the queue back, turn off Steer agents while they work in Settings > Agents, or select Queue for one agent in its settings. If you already changed this setting, your selection stays. An agent whose provider cannot steer still queues the message. A joined server keeps its own default until it is on the new version.
A request from another agent now waits for a turn of its own while the recipient works, also when the recipient steers. Before, a recipient with Steer selected took the request into its running turn, and the result for the other agent and the answer to you became one message. An answer or a notice from another agent still joins the running turn. Routine runs, channel work and messages from Slack, Telegram and WhatsApp start their own turn, as before.
A steered message no longer makes an agent drop the work that it does. OpenBot now tells Codex, Pi, Muse and the models of an API key provider that the message adds to the current work: the agent answers a question first, applies a correction from that point, and then continues. Claude Code and Grok Build already did this.
Settings pages and the full browser view have a clear Close (X) button at the top right. On a settings page it goes back to the chats, as Escape does. In the full browser view it replaces the small "minimize" icon and puts the browser back in the side panel.
On a joined server, a member who can only view an agent sees its Library folders but cannot change them. Owners and admins can change all folders.
A hosted server that runs on one of our new machines has more storage: 25 GB on Starter, 75 GB on Standard and 150 GB on Pro. Settings, Billing and the plan change offer show the storage of your server. For the agents of the server to know the new storage, the server must also be on the new version.
The models of a custom provider on this computer are now a table in a wider dialog: one line for each model, with a name above each column. The efforts and the input kinds of a model are menus, and a token count shows in groups of three digits, such as 1,048,576.
A new OpenCode agent starts on Beam Free, the fastest free model in our test. The model picker shows Beam Free, Big Pickle and GLM-5.3-Flash Free first in the OpenCode group, and no longer folds them under Older models.
In the web app, the loading screen with the falling plushies now stays until the agents, their order, the unread counts and the channels are ready, so you do not see the sidebar load. The screen goes after 5 seconds at most when the server is slow to send the sidebar. Chat messages can still load after the screen goes.
A hosted server that runs on a dedicated machine keeps its data for 30 days after its plan ends. Then OpenBot deletes the server and its data. Renew the plan in these 30 days to keep the server. A plan whose payment is still retried keeps the server.
A server with no logo of its own now shows a felt plush in one of the agent colours, in place of the colour gradient or the initials. It shows in the server list, the server menu, Billing, Notifications, and in Server Settings for a member who cannot change the logo. It also shows in the web client.
The privacy notice now says that an operator can move a Standard or Pro hosted server to a Leaseweb virtual server in Amsterdam, and that the Leaseweb reference of that server holds the account email and the first 8 characters of the server ID.
A muted agent stays in its folder when it has unread replies. It does not go to the Unread group, and "Unread first in folders" does not put it first. Its row keeps the unread count. Unmute the agent to move its unread replies into Unread again. This also works on a joined server.
Buy a cloud server on the same pricing page as the website. Add a server opens the page in the app window, with the server, the license and the price in one form. "Your computer" replaces "No server", and a license alone goes to its own payment page. This also works in the web app.
The app and the web app can show and charge server and license prices in Brazilian reais (BRL) when the account server offers them. A computer in a Brazil time zone gets BRL first.
The Team license in the order form, on the website and in the app, starts at 2 members.
A generated image now has a cleaner frame. The status, the size and the prompt show under the image, and the download and preview buttons show on the image when you move the pointer over it. A failed or stopped generation shows a smaller frame with the reason under it.
When we move your hosted server to one of our new machines, the move now also keeps what you installed on the server: programs in /opt and /usr/local, npm packages, apt packages, your crontab, and the agent CLIs that you signed in to. Your provider sign-ins, API keys and MCP servers move with your files, as before.
A moved server is offline for a shorter time: we install your apt packages before the move, and copy your files straight to the new machine.

Fixed

On a server, the browser sign-in view now opens after a saved login did not work. Before, it said that the browser view was protected, and you could not sign in by hand. The page loads again first, so you start the sign-in from its first step. The server must also be on the new version.
The Library and the Storage tab of a joined server show a member only the files of the agents that the member can see. A file that the server cannot connect to an agent is now hidden from a member who has a closed agent. The server must also be on the new version.
When Google asks you to verify your account, the Gemini error banner again shows a clear message and a Verify account button that opens Google's page. Before, with the new Gemini runtime, the banner showed Google's text with the verification link cut off, and the link did not open. (#6985)
ChatGPT agents no longer show a "Provider error" message when Codex cannot update a plugin marketplace at start, for example on Windows without Git. The chat continues to work as before.
Show a notification with the progress while a provider downloads for the first time, on this computer and on a joined server. Before, only the provider row showed the progress, and on a server the row could stay at "Not downloaded" until the download ended. The notification also shows for Hermes, and for the download of Pi that starts when you save a key for Inception, DeepSeek or OpenRouter.
Show only the download on the settings page of a provider that is not downloaded. Before, the page also showed Runtime, Models and Use, which had no effect yet.
An agent that asks for the models of one provider now gets the reason when that provider has none, for example that Gemini is off or not downloaded. Before, the agent got an empty list and could not tell you what to do.
An agent can now read the next page of an older chat when it reads that chat from another chat. Before, the agent often got "History is unavailable for this request" and could not check earlier work. When the agent uses the page link in a search, it now gets the reason and the correct step.
Sign in to an MCP server again after the server stopped accepting your sign-in. Sign in now opens the browser. Before, Sign in on the card in the chat or in Server settings, MCP, opened no browser and showed "The server did not accept that sign-in." again. On a joined server, the server must also have this version.
When an agent installs one skill for many agents, the chat now shows one row with the number of agents, for example "Installed skill Release notes · 38 agents". Before, it showed one equal row for each agent.
An agent that revises a skill now knows that it must not edit the installed copies of the skill. An edited copy stops later updates of that skill.
Apple apps can open an email in Apple Mail again. Before, list and search worked, but reading, marking or deleting an email always failed with "Can't get object. (-1728)". Apple apps that you already installed get the fix when OpenBot starts. On a server, the server must also be on the new version.
The model picker now shows the model when a provider has more than one account. Before, a long account email took the room, and the name was cut to "Claud…". The picker now shows "Opus 4.5" and not "Claude Opus 4.5" there, because the provider mark is beside it.
Keep the agent's latest message in the sidebar while a routine runs in the agent's chat. Before, the sidebar showed the routine's task as a new message, and it kept that text when the run failed or when OpenBot closed during the run, while the chat did not show that message.
When Claude cannot read its sign-in, for example while the macOS login keychain is locked, an agent that switches to backup models without asking now moves to its next ready backup model. An agent with no backup model now shows "Your Claude sign-in expired. Sign in again to continue." Before, each Claude turn failed with "Not logged in · Please run /login", and no backup model was tried. On a joined server, the server must also be on the new version.
In a channel, a task now completes when its member posted the result and the model reply then failed, for example with "The model's reply was cut off or was not valid." Before, the result was in the chat, but the task showed as failed, and the agent that gave the task waited for it until you sent a new message. On a joined server, the server must also be on the new version.
On Windows, you can now delete an agent that used OpenCode. Before, the delete failed with "The agent data could not be removed completely" each time, and the agent came back, until you restarted OpenBot. If a reply of another OpenCode agent is in progress, try the delete again after the reply ends. On a server, the server must also be on the new version.
When an MCP server from your own Gemini settings does not start, the chat now names the server and tells you to remove or repair it, then start a new chat. Before, on Windows, the chat often showed only "The agent could not continue. Try again." On a joined server, the server must also have this version.
Read the plan usage less often while a usage limit holds agents. Before, each check of a held limit read the usage again, also right after the provider refused a read, so the usage in the sidebar could stay out of date for an hour.
An OpenCode chat now compacts its context automatically at 80%, and the context meter has "Compact now". Before, the meter showed "This agent does not compact its context.", and a long chat sent an almost full context with each message. This used more tokens than necessary. On a joined server, the server must also be on the new version.
When OpenCode refuses a free model, for example because the model is not available in the user's country, the error now says so and tells the user to choose a different free model. Before, the error said that OpenCode had no sign-in and told the user to choose a model with the Free label, although the chosen model had that label.
More sites with a bot check now load in the agent browser. The browser sent a version text and a language header that no Chrome sends, and some sites refused it.
The agent browser now waits for a Cloudflare check on a page that also shows a cookie banner. Before, it told the agent that the check had passed while the check still ran.
The agent browser now gives a site the browser details that the site asks for, as Chrome does.
The agent now gets the page from its first read when a site redirects immediately. Before, it had to read the page a second time.
An agent on a custom endpoint, such as a local model, is no longer stopped with "stopped responding" after 5 minutes of silence. A custom endpoint now gets 30 minutes before OpenBot ends the turn. On a joined server, the server must also be on the new version.
Escape in the message input keeps the text cursor. Before, after Escape closed the @ list, Backspace did not delete the @.
The input of the diagram chat grows with a draft of more than one line. Before, the text went out of the input.
An agent that reads older messages of a chat by a message ID now gets them. Before, it got "History is unavailable for this request", and a routine that reviews earlier work could skip messages. When the message is not in that chat, the agent now gets an error that tells it what to send instead.
A muted agent now also plays no sound. Before, only its notifications stopped.
End a usage limit hold soon after the usage is reset outside OpenBot. Before, an agent set to switch without asking moved each new message to its backup model and did not try its own model, until the reset time or a restart. Now a message tries the agent's own model first. OpenBot also reads the usage again every 15 minutes during a hold, also when the provider named the reset time.
An agent that looks for its WhatsApp tools now finds the list of connected WhatsApp numbers and chats. Before, an agent could say that no WhatsApp connection was available, also when the number was linked.
Install a new revision of a local skill when the installed copy has no unsaved edits. Before, the install stopped with "This skill has local changes" when the folder had a macOS .DS_Store file, when the agent saved the revision from its installed copy and the description had a colon, or when the files matched a newer saved revision. A copy with real edits is still kept until you replace it. For agents on a joined server, the server must also be on the new version.
A Grok agent at Workspace only now uses OpenBot's own tools, such as sending a message to another agent, without an approval card. Before, each of these tool calls waited for Allow, so a scheduled run stopped when nobody was there. A write outside the workspace and a tool of another server still ask. On a joined server, the server must also be on the new version.
When the GitHub connection or Computer Use is on, an agent now finds it in its list of MCP servers. Before, the agent's setup showed the github server, but its server list did not, and the agent said that no server had that name. The agent now says that the server is part of OpenBot and where you turn it on or off. A joined server must also be on the new version.
OpenCode agents now find the tools of a connected app with many tools, such as n8n. Before, the agent said that some of these tools were not enabled, but the server test showed that they were ready.
OpenBot starts faster, most of all on Windows: it reads its settings files together, it copies its skills to the agents together, and it no longer waits for the removal of old provider CLI versions before the window shows.
A slow computer stays usable after OpenBot starts. OpenBot started all provider CLIs in the same second. Now it starts the provider of your chats first, so chat is ready sooner, and then the others two at a time.
OpenBot no longer writes its skill files again at each start when they did not change. An antivirus program does not scan them again, and a provider does not load them again.
A Codex agent now asks you in OpenBot before an action that needs approval, also when auto review is on in your Codex settings. Before, the Codex auto reviewer made the decision. It did not accept your answer in an OpenBot card, so the agent could not continue after you approved.
Keep the pictures of a joined server's agents while the server is offline. Before, during a restart of the server for an update, each agent showed its generated avatar until the server came back, and the picture looked lost. The desktop app and the web client now keep a copy of each picture that they loaded; sign-out and server removal delete it.
When Gemini refuses a request with "Request contains an invalid argument.", the chat now tells you to choose another model and send your message again, or to start a new chat. Before, the chat stopped with only Google's text. On a joined server, the server must also be on the new version.
When the processor of a computer cannot run a provider, such as Gemini (Antigravity) on a processor without AES or PCLMUL, OpenBot now says that the processor cannot run this provider and that you must use a different provider. Before, the row said only "Try again", and each new try failed in the same way. On a joined server, the server must also be on the new version.
When you gave an agent a new key for an installed app, such as Obsidian, the agent saved it in the Vault and said that the app was fixed. The app did not use that key and kept its old one. The agent now tells you to open the app in Marketplace and click "Change key".
On macOS, Computer Use can now read a file or folder picker that opens on a window, for example Add folder in an agent's settings. Before, the agent could not read the picker, and you had to pick the folder by hand. OpenBot now ships version 0.41.3 of the Computer Use driver. A joined server must also be on the new version.
A Gemini (Antigravity) agent can now use all tools of an MCP server from OpenBot when your own Antigravity settings have a server with the same name. Before, Antigravity used its own entry, and the agent could not call a tool that this entry did not turn on.
A Marketplace app that starts a local server with npx (Obsidian, Airtable, Apple apps, Brave Search, Firecrawl, Resend) now starts with the runtime that OpenBot downloads, not with the Node.js of the computer. Before, Obsidian showed "The server stopped before it answered" on a computer with Node.js older than 24, and the agents could not use it. The first connection can wait for the download of the runtime. A server that you added yourself starts as before. For an app on a joined server, the server must also be on the new version.
When Claude says "You're out of usage credits", an agent that switches to backup models without asking now moves to its next ready backup model. Before, the turn failed with this text, and no backup model was tried. An agent with no backup model still shows the text. On a joined server, the server must also be on the new version.
When an OpenBot program file is damaged or missing, OpenBot now tells you to install it again and gives a button for the download page. Your data stays. Before, it showed only "A JavaScript error occurred in the main process" and did not open.
With the transparent window on, the controls that hang over other content are solid again: the mention list, the search bar of a chat, the button that goes to the latest message, the subagents bar, the drop areas, the Library selection bar, the bars of the workflow canvas, the nodes of a diagram, the loading screen and the status badges. Before, the content under them showed through and their text was hard to read.
Keep the files that an agent puts in OpenBot's API key folder. Before, each Vault change and each start of OpenBot deleted every file in that folder that was not a Vault key, such as a private key file that cannot be downloaded again.
When an agent asks for a new chat and only routine runs wait in its queue, the new chat now starts before those runs. Before, the first routine run cancelled the request. An agent with frequent routines almost never got its new chat, and its chat kept growing. The runs stay above the "new chat" line, so the agent does not see their answers. A message from you or a teammate that waits still ends the request, as before.
Routines now run on time after the computer clock is corrected. Before, when a computer started with a wrong clock (for example after a power cut) and the clock was corrected later, no routine ran for hours, until you saved a routine or restarted OpenBot.
In the workflow editor, holding Space now moves the board after you click a button on a card, such as "Trigger type" or "Run". Before, Space opened the menu of that button again, or ran the step, and the board did not move until you right-clicked. A button that you go to with the Tab key still uses Space.
On Linux and Windows, OpenBot now completes its shutdown when you close the main window. Before, the process could end in the middle of the shutdown.
When GitHub is connected but refuses a write, such as a push or a new repository, the agent now tells you how to correct it: in Marketplace > Apps > GitHub, select Choose repositories and install OpenBotGit on the repository. Before, the agent asked you to find a GitHub card that did not show, also in channel chats. A joined server must also be on the new version.
In the web client, an agent reply to the open chat no longer moves the "New" line to the reply. Before, each reply showed a new line and the chat jumped. A reply that comes while you are away from the tab still gets the line.
A scheduled channel routine with nothing new to report now ends quietly: the channel gets no message and no notification. Before, the agent got the error "Use this tool only in a routine run", and it often posted a "no change" message in the channel. A Test run still shows the answer, and a member that works on a part of the run gives its result to the owner of the request, as before.
An OpenCode chat now gets the OpenBot instructions, memories and teammate list one time, also when you write again after a pause or after you use a different agent. Before, each of these added one more full copy to the chat, so a long chat used many more tokens than necessary. On a joined server, the server must also be on the new version.
Agents in the browser now see and use a visible form that has a field named "hidden" or "inert". Before, such a form did not show for the agent, and the agent could not click or fill its fields, but the form showed in screenshots. On a joined server, the server must also be on the new version.
An image that an agent makes now shows one time in the chat. Before, ChatGPT agents often attached a copy of the image to the answer, so the same image showed two times. On a joined server, the server must also be on the new version.
On a joined server, a member no longer gets the stored files of a deleted agent in Storage. Owners and admins still get them.
When an agent asks for a local skill or a skill revision that does not exist, it gets a clear message. The message says that the skill does not exist, or gives the latest revision. Before, the agent got a system error with a file path (ENOENT). It then could not correct its call, and the install of a skill for another agent stopped.
The Restart server action no longer shows for a hosted server that it cannot restart. Before, the action showed and then failed.
Find models on a custom provider now fills the context window and the max output of a model when the endpoint lists only the model name. OpenBot reads the values from the model catalog that OpenCode keeps on the computer or the server. Before, such a model got 128,000 tokens until you typed its context window, so a model with a larger window compacted its chat too early. A value that the endpoint publishes, or that you typed, stays as it is.
A server on this computer that did not start when OpenBot opened, or that stopped during its start, now starts again by itself, a few times. Before, it stayed stopped until you selected "Restart server" in its menu. When it still does not start, Retry on its screen now starts it. Before, Retry only tried to connect again and could not help.
The Gemini usage ring in the sidebar shows your Gemini limit again. Since 0.67.0, Google did not send the Gemini quota to OpenBot, so the ring was missing or showed the Claude and GPT limit. On a server, the server must also be on this version.
A channel task now shows the result of a Claude member that finishes after a long background job, such as an install or a dev server start. Before, the task closed when the member's turn ended, and the result that came after the job stayed only in the agent's own chat.
Discord, Slack and Telegram now connect on a server that you made with "New server". Before, Discord and Slack showed "Sign in is required." while you were signed in, and Telegram stopped after 15 seconds with "The operation was aborted due to timeout".
Keep the first Gemini account in the usage popover while Gemini is idle. Before, the account went from the list about an hour after its last session, and only the accounts after the first stayed.
Say on the first Gemini account when a saved Gemini API key is used in its place. The key does not give Gemini 4 Carbon or the Claude and GPT models of the Google plan. Before, the account kept a Reconnect that opened nothing, and the models changed with no reason shown. Remove the key to use the account's plan again.
Show the Google account email on each Gemini account. Before, every account showed only its label, such as "Account 1", so two accounts could not be told apart.
Replies in Italian are read aloud with the Italian voice. Before, a reply with "è", "più" or "città" was read with the French voice.
OpenBot stays responsive after a restart when you have many tasks or threads. Before, the window and the connected apps were very slow for several minutes while OpenBot opened its chats again.
Let Gemini agents (Antigravity) answer when an MCP server has a tool whose input format Gemini cannot read, such as the email of the Resend server. OpenBot now leaves that server out of the chat, tells you which server and tool it left out, and sends the message again. Before, each message stopped with "Agent execution terminated due to error.", and the chat did not name the server.
Connect the TradingView app, and other MCP servers whose firewall blocks sign-in pages that name 127.0.0.1 or localhost. When the browser does not come back on either address, try the sign-in again: OpenBot then returns through the openbot:// link, and the browser can ask to open OpenBot. Before, every attempt opened a blocked page.
The chat card of a new server shows its logo, its name, its status and "Open server" on one row again, like the other chat cards.
When you open a chat with an agent that waits for your answer or approval, the chat no longer jumps. The question card and the "Waiting for your answer" line show with the chat.
Text that you copy from a chat no longer has empty lines at its start or end. Before, a triple-click or a long drag copied the text with empty lines, so a paste into the message box or a terminal did not stay on one line.
Claude agents use less of your plan when your Claude settings set ENABLE_TOOL_SEARCH to auto and the model has a 1M context. Before, Claude sent the full definition of each tool in each request. Now OpenBot keeps tool search on, and Claude loads a tool only when it needs it.
A Claude agent that continues a chat after a pause waits for its MCP servers before its first request, so that a slow server does not seem to have disconnected.
ChatGPT agents now work when Codex on your computer is set to use a different model provider, such as a local gateway. Before, OpenBot asked you to connect ChatGPT again and again, and the connection failed with "OpenBot could not verify the ChatGPT connection". When you connect ChatGPT or add an OpenAI API key in this setup, OpenBot now tells you that Codex uses the other provider.
Switch a Claude agent to its backup model when Claude refuses the subscription for a billing problem, such as a payment that is past due. Before, the agent stopped with the error and did not try the backup model, also when automatic switching was on.
Show only the sub-tasks of the current round in a channel's "Waiting for sub-tasks" list. Before, when the lead agent gave out work again, the list also showed every sub-task of the earlier rounds as replied and not read yet, so the work looked stuck. For a joined server, the server must also be on this version.
Show the agent cursor of Computer Use where the agent clicks or scrolls in a window. Before, it could show on the main display, or away from the window, while the agent worked in a window on another display or a window that was not at the top-left corner of the screen.
After a move to a new machine, Pi agents no longer time out at their start, and the server starts about 2 minutes faster. The move removes the settings of the old remote desktop, and the server no longer waits for that desktop.
When you take over an agent's browser tab for a security check, select Done, and the check is still there, the agent now tells you that the site blocks its browser. Before, the agent asked you to take over or to sign in again, and each new try ended on the same check. For an agent on a joined server, the server must also be on the new version.
An agent can now turn off an MCP server that is running. Before, the agent tested the server first, and a server that lets only one copy run failed the test, so it stayed on. A server that is off is now saved without a test.
On a slow or unstable network, a joined server no longer shows "The remote request timed out." and fails to load the open chat while the server still answers. Before, the app stopped waiting for some reads after 15 seconds, also when they only waited behind earlier answers. Now a read fails only when the server sends nothing for 15 seconds, or after one minute.

Nov 11, 2026

Yaji 0.68.0

New

Channel settings have a Browser setting. Choose if every member, one member or no member can use the browser in the channel. The other members give their browser steps to the browser member. On a joined server, the server must also be on the new version.
The channel side panel has a Browser pane and a Files pane. The Browser pane shows the tabs of the channel's members, and a new tab opens for the browser member. The Files pane lists the files sent in the channel. On a joined server, the server must also be on the new version.
When a new version of OpenBot cannot finish its start 3 times in a row, the next start shows a small recovery window. You can try again, download the previous version, export diagnostics, or open the data folder. Before, the app did not open and you did not know why. A quit, a shutdown of the computer, or a stop from Activity Monitor during the start does not count as a failure.
You can turn off the "Unread" group in the sidebar. Right-click the "Unread" heading and select "Turn off Unread group", or set "Sort chats" to "Manual" in Settings > General > Conversations. The web client has the same setting in Settings > Preferences.
Right-click the Dynamic Island or the agent dock to hide it. The menu item has the same effect as the switch in Settings, and you can show it again there.
Drag an agent or a channel on the dock to put it in a new place.
Drag the Dock size slider in Settings > Agent dock to make the dock smaller or larger, from 70% to 140%.

Improved

A new desktop version now goes to about 1 in 10 computers first, and to all computers after we check its crash and start numbers, usually one day later. While your computer waits, Settings > Updates says that the new version is rolling out. To get it now, click Check for updates. Servers and Docker images still get each version when it is published.
Agents now ask questions with options in the question prompt. Cards with buttons are only for an approval, such as of a message draft. This makes each agent request about 500 tokens smaller.
Before we publish a desktop release, it must now open the data of the previous release on macOS, Windows and Linux. The check installs the previous release, makes agents, chats, memories and a channel, installs the new release over it and makes sure that all of them are still there and that the database is complete. A release that cannot open your data stops before it is published.

Fixed

When OpenBot quits or restarts, Hermes agents with Computer Use no longer lose a turn that starts at that time. Before, OpenBot stopped Computer Use before it stopped the agents, so a routine or a message that started then failed before the agent could answer.
On macOS, a sign-in in the built-in browser no longer stops when the site asks for a passkey. The built-in browser cannot show a passkey prompt, so the page waited, with nothing on screen, until the site's time limit. Now the site shows its other sign-in methods at once, for example the Microsoft Authenticator app or a password. When the site asks for a security key, or for a credential that a security key can hold, the request does not change.
On Windows, the agent dock and the Dynamic Island no longer make the whole computer slow while many agents work. Before, each mouse move on the computer waited for OpenBot while the dock or the island was on screen, so the pointer lagged in all apps. Now OpenBot follows the pointer only when it is near the dock or on the island.
On Linux, agents can use the browser while the OpenBot window is hidden or minimized, and in the chats of a local server. Before, pages in those tabs stopped drawing: screenshots timed out, dialogs that open with an animation did not appear, and when you took over a tab, it showed only "Connecting to the page on the server…". On a joined Linux server, the server must also be on this version.
An agent can again create a channel, send feedback, or do other actions that only your request can start, when you ask for it in its chat while it waits for answers from teammates. Before, the answers that came in early went into the same turn as your message, so the agent refused your request and asked you to send it in the chat. Now your message starts its own turn, and the agent reads the teammate answers together when the last teammate answers.
When OpenBot cannot send an agent's bug report or idea, the agent no longer opens the issue on GitHub by itself. Before, an agent with GitHub tools wrote the issue without the OpenBot issue form, and it was closed at once. Now the agent describes the problem in its reply and gives you the link to the issue forms. On a joined server, the server must also be on the new version.
After a usage limit, a provider error or a refused session, the answer to the messages that waited now shows below the last of them. Before, when the waiting messages ran later, for example on another Claude account, the answer showed below the first message that hit the limit. The messages sent after it looked like they had no answer.
Computer Use no longer draws its blue frame around an app that the agent does not use. Before, when the agent worked in an OpenBot window or in a window on another Space, the frame went around the window in front, often the app that you used. Now no frame shows until the agent works in a window that is on your screen.
Agents no longer say that sidebar folders are flat. They now know that a folder can hold folders, and that when you delete a folder, its agents and folders move up one level. On a joined server, the server must also have this version.
A webhook routine on a hosted server that sleeps now starts the server. Before, the server stayed asleep and answered each delivery with 503, so the routine did not run until someone connected to the server. The first delivery still gets 503. A later retry, after the server starts (usually a few minutes), reaches it.
An agent can now open a new page in its browser while the current page still loads, as pages of Next.js and other streaming sites do. Before, the agent got "Navigation failed (-3)" and the tab stayed on a half-loaded page with no controls.
The steps of an answer from a ChatGPT (Codex) agent show its thinking. Before, Codex sent no reasoning text, so the steps had no "Thinking" step. Your own model_reasoning_summary setting in the Codex configuration still applies. On a joined server, the server must also be on this version.
A Claude agent now gets the tools of an MCP app, such as Composio, from the first turn after you enter its API key. Before, the agent could have none of its tools for up to 15 minutes, and its session reported that the app needs a sign-in.
When a Grok model answers with only blank text, the chat now says that Grok ended the turn with no answer, and tells you to send the message again. Before, the turn ended as done, and the chat showed no reply and no error.
A form that an agent shows in the chat no longer fails the first time when the agent adds a preview to it. Before, OpenBot told the agent that every card type can have a preview, but only a confirmation card can. The agent then had to try again without the preview.
A Claude agent now answers the result of a background subagent when the subagent finishes after the agent's turn ended. Before, if another thread or another Claude agent finished a turn at the same moment, OpenBot could close the agent's session first. The result then stayed unread until you sent a new message, and the agent's first reply was "No response requested."
OpenCode agents now use browser scripts correctly more often. Before, an OpenCode agent sometimes put the browser steps of a script, such as opening a page, directly in its own code, where they do not work. Each of these steps failed, and the agent then did the same work one step at a time, which took longer. The agent now gets a note that these steps work only inside the script.
The OpenRouter model list no longer shows the "(batch)" models. They work only for batch jobs, so a chat with one of them failed with a 404 error.
On a server, the model picker of a second Gemini (or other provider) account now shows the models of that account's plan, such as Gemini 4 Argon on an Ultra plan. Before, it showed only the models of the first account. The server must also be on the new version of OpenBot.
The Gemini usage ring shows again on a plan for which Google sends no Gemini limit. It shows the "Claude and GPT models" limit, and the usage popover names it.
An agent no longer stays on "working" with no end when a browser tab stops. This can occur when the page shows a dialog during an action. After two minutes, the agent's browser call returns an error, and the agent can tell you to answer the dialog or open the page again.
When OpenAI refuses to start live voice with a 403 error, the voice panel now says so in plain words: try again later, or turn off live voice in Settings to talk by voice. Before, it showed the raw Codex error "unexpected status 403 Forbidden: realtime websocket handshake failed". The technical details stay in the log.
An agent can search and read its earlier chats for the full reply. Before, when a memory review ended or you changed the agent's memories or settings while the agent answered, each history search and read in the rest of that reply failed with "History is unavailable for this request".
An agent set to switch to its backup models automatically now switches when its model, such as Grok, gives only its reasoning and no answer before the agent starts to work. The message then runs again on the first available backup model. Before, the turn failed, and the backup models were not used. A turn that already used a tool still does not run again, so a tool does not run two times (#6594).
When Computer Use is on for an agent but not ready on this computer, the agent now says so and tells you to open Settings, then Computer Use, to start it or see what is missing. Before, the agent looked for Computer Use tools that it did not have and could not tell you why.
An agent that saves a channel memory longer than 500 characters now gets the limit as the reason, and it can save the facts as shorter memories. Before, the save failed with "Provide the memory text.", and the agent tried the same text again.
Agents such as Kimi now reach an MCP server on a network or with an antivirus that checks secure connections. Before, the server test passed, but each tool call failed with a certificate error. Agents now trust the same certificates as your computer. On a joined server, the server must also be on this version.
Start a new Claude Code session with the chat history when a chat or channel gets too long for Claude. Before, each later message failed with "Prompt is too long", also after a restart, until you started a new chat. Now the error tells you that OpenBot started a new session, and your next message works. In a channel, click Continue on the task. On a joined server, the server must also be on this version.
Stop a channel task when OpenBot replaces the member's provider session after an error. Before, the task showed the agent as working and blocked the member's next tasks in that channel.
A webhook routine now works with senders that use the Standard Webhooks format, such as PostHog, when you give them the routine's secret. Before, the sender refused the secret as not valid base64, and when you converted it, the routine refused each request with 401. New secrets are now in the format that these senders need. A secret that you made before still works when you convert it to base64 in the sender, or you can make a new secret in the routine editor. For a routine on a joined server, the server must also have this version.
A message that an agent sends to a Telegram group or a Discord server now shows the name of the group, the server or the bot as written. Before, each of these names, when it had 8 or more characters, arrived as "[redacted]", for example "Setor: [redacted]" in a group with the name of the sector.
A channel now answers new messages after a member's background job ends. Before, when a Claude member answered a finished background job by itself, the messages that the channel sent to that member while it answered did not start until OpenBot was restarted. The server must also be on the new version of OpenBot.
When the model service of a Pi agent refuses a request, the chat now says that the model service refused it, shows the reason of the service, and tells you to select a different model or check your access to the model at the service. Before, the chat showed "You do not have permission to complete this action. Ask the owner for access.", which pointed to OpenBot and hid the reason. On a joined server, the server must also be on the new version.
When a usage limit stops an agent after it started work, and no backup model is ready, the work now waits in the queue and continues when the limit ends. The agent continues from the work it saved and does not do the completed steps again. Before, the message failed, but the agent showed "Waits for limit", and nothing ran at the reset. On a joined server, the server must also be on this version.
When another Codex app, such as the Codex desktop app, has a ChatGPT agent's chat open, a message to that chat now says that the chat is open in another Codex app and that you must close it there, then send the message again. Before, each message failed with the technical text "already has an active writer", and a restart of OpenBot did not help.
A channel routine run now shows as failed when its work could not be done. The agent that owns the request, also a member that the lead gave the request to, can end the run as failed with a reason, and the run history shows that reason. Before, the agent got "Use this tool only in a routine run.", and the run showed as succeeded. For a channel on a joined server, the server must also be on this version.
When "Steer agents while they work" is on, a message that you send to a Claude agent as it finishes its answer now shows below that answer and above its own answer. Before, the message showed above the earlier answer, and the two answers showed one after the other below your messages.
A Claude agent that moves to a backup Claude model after a provider error, such as a network failure, now answers the message on the backup model. Before, the agent stayed "working" with no answer, and the messages behind it did not start until a Stop or a restart.
When a channel task cannot start because the agent's provider is signed out, the stopped task now says what to do, in your language: for example, that your ChatGPT sign-in expired and that you must sign in again, or that the provider needs an API key. Before, the task showed only "The provider could not start this assignment. Resume to try again." in English, and the reason was only in the log. After you sign in again, click Continue on the task. On a joined server, the server must also be on this version.
When a step of a browser script fails because of an internal error in OpenBot, the agent now gets the error, the number of steps that ran, and a snapshot of the page. Before, a ChatGPT agent saw only "dynamic tool request failed", also when a click or selection had already run, and could do it again.
A chat no longer stops working after a model refuses it because the chat has too many images, for example a Venice model in OpenCode after a tool returned more than 10 images. OpenBot now starts a new session with the chat history, and the agent continues on your next message. When the model names the image limit, OpenBot tells you that the chat has too many images. Before, each new message failed again with "Invalid request parameters".
A Grok agent now finds its OpenBot tools, such as the tool that sends messages to teammates. Before, Grok showed these tools only through its tool search, and an agent sometimes said that team messaging was not available and sent nothing. A joined server must also be on this version.
When a Kimi model step fails and Kimi ends the turn with no answer, the chat now says that Kimi ended the turn with no answer, and tells you to send the message again. An agent set to switch to its backup models now uses them. Before, the agent went back to ready, and the chat showed no reply and no error (#6722). On a joined server, the server must also be on this version.
Agents can read online videos on a computer that connects to the internet through a proxy. Before, OpenBot downloaded its video tool without the computer's proxy settings, so the download failed and the agent got "Check the internet connection", but the built-in browser opened the same link. The download now uses the same proxy settings as the built-in browser.
Agents can click elements from their newest browser snapshot on pages that load content in frames, such as Gmail. Before, a frame that loaded in the background made each click fail with "Stale browser reference", also after a new snapshot.
On Windows, installing or updating a skill for an agent no longer fails when antivirus or the search indexer opens the new skill files for a moment. Before, the install stopped with an "operation not permitted" error, and a temporary .openbot-stage- folder stayed in the agent's skills folder, where the agent could load it as a second copy of the skill. Now OpenBot waits up to about 3 seconds and tries again. When the install still fails, it also tries to delete the temporary folder.
An agent that asks for a routine with a wrong schedule now gets told which part of the schedule to correct, so it can create the routine. Before, every wrong schedule got only "The routine schedule is invalid.", and some agents, often on OpenCode, tried many formats and then said that they could not schedule the work. For an agent on a server, the server must also be on the new version of OpenBot.
A workflow that an agent builds to post into an OpenBot channel now posts there. Before, the agent chose the wrong output type, and each run stopped at the output step with "Choose a messaging place for this output." To repair a workflow that an agent made before this fix, open it, click its output step, and choose the channel again.
An agent set to "Ask first" that creates a new agent now gives it "Ask first" too. Before, the new agent started with "Full access" and acted without asking. On a server, the server must also be on the new version of OpenBot.
An agent that answers on WhatsApp now sees the messages and files that OpenBot sent to that chat, also when another chat, a teammate or a routine sent them. Before, the agent said that it could not see the message, and you had to send it again. On a joined server, the server must also be on this version.
An image that OpenBot sends to WhatsApp now shows in its full shape in the chat. Before, the chat showed it cropped until you opened it. On a joined server, the server must also be on this version.
When an agent cannot publish, replace or delete an openbot.site site, it now gets the reason, for example a source that is not a folder or an unknown site, and can correct the call. Before, a ChatGPT agent saw only "dynamic tool request failed" and tried the same call again.
When an agent sends a message to a teammate, or stops a teammate's work, with an agent id that has a wrong character, the error now names the teammate with an almost equal id, so the agent can try again. Before, the agent got only "Unknown OpenBot agent" and could tell you that the work was blocked, also when the teammate was ready.
A Gemini agent that cannot open a file in another folder now tells you which folder to add in its settings, under Folders. Before, a Gemini agent in Full access was told that it could open every folder, so when Gemini refused a project folder, the agent said that no setting could add it. On a server, the server must also be on the new version of OpenBot.
Pi starts and works when one of your Pi extensions or packages writes text to its output. Before, OpenBot stopped Pi with "Pi returned an invalid RPC record.", and you could not add a Pi agent.
A ChatGPT (Codex) agent no longer shows a provider error when the Codex settings ask for a faster speed that the model does not have. Before, the chat showed "Configured service tier priority is not advertised as supported" each time the agent continued after a pause, but the message went to the agent as usual at the standard speed.
The Usage popup now shows current plan limits for Claude, Codex and Grok also when no agent used them for a while. Before, after OpenBot stopped an idle provider, the popup kept its last reading until a limit window reset, sometimes for days, so use outside OpenBot (for example in a terminal) did not show. On a joined server, the server must also be on the new version.
Sign in to an MCP server whose sign-in server gives plain OAuth information at its OpenID address, such as the TikTok organic connector. Before, the sign-in stopped before the login page opened, and the server card showed invalid_type errors for jwks_uri and two other fields.
GPT live voice now speaks the language that you speak, also for its own short words such as greetings and "Understood.". Before, the voice said these words in English, even when you and your agent used a different language. For a joined server, the server must also be on the new version of OpenBot.
When a message waited for the sign-in of the agent's provider and you selected a model of another provider, the change did not apply. The message waited with no end, and the chat asked you to sign in to the old provider. Now the new model applies at once and the message sends with it. On a joined server, the server must also be on this version.
A Codex chat that stays almost full after a compaction no longer compacts again after each turn. Before, Codex reported a small estimate while it compacted, so OpenBot did not see that the compaction freed no space. The chat compacted without end, with no Stop button and no message. Now OpenBot stops after 2 such compactions and tells you to start a new chat.
When an OpenCode model's reply is cut off or is not valid, the chat now says so in plain words and tells you to send your message again or choose another model. Before, it showed only the provider's internal error text, such as "Internal error: OpenAI Chat stream ended without finish_reason".
When the desktop app reaches a joined server through Signal but the direct connection does not come up in time, it now says that a firewall, VPN or proxy can block the connection. Before, it said only "The host is not reachable.", also when the server showed as online. Other failed connections still show "The host is not reachable.", now with a code that names the failed step. The app still tries again by itself.
When an agent sends a bug report or an idea after your account has sent too many today or in the last minute, the agent now tells you why OpenBot did not send it and when it can send again. It also gives you the link to submit the report on GitHub. Before, the agent said only that OpenBot could not send it, and it often tried again and failed again.
A file change that you approve for a Gemini agent now shows as done in the chat. Before, the chat also showed a failed step, "Tool call was approved but never executed.", beside each approved change that worked, and an agent that read the chat later counted these steps as errors.
An agent can upload the same file in many browser tabs, for example one video to several social sites. Before, each upload kept its own copy of the file until its tab closed, and after about 500 MB the browser refused all new uploads. Now an upload of an unchanged file uses the copy that OpenBot already keeps. When the limit stops an upload, the agent is told to close the tabs whose uploads are finished.
Moving agents or sending a large file to a joined server on a slow connection now finishes. Before, when one part of the file did not send, the retry could fail with "The WebRTC file offset is invalid." and the transfer stopped. The server must also run this version.
An agent page in the chat now lines up with the agent's messages. In a channel, the page starts after the column of agent faces, as the agent's bubbles do. On a wide window, the page is as wide as a wide agent bubble, not 720 px.
When the web client opens a server, the page no longer moves while it loads. Before, a second loading screen showed after the saved agent list, the sidebar moved agents into the Unread group, and an image in the open chat grew from a small file card to its full size.
Settings now open as one complete page. Before, the page showed empty for a moment on the web, and the Profile blocks faded in one after another.
The Agents rows of Server Settings no longer move when their values arrive from the server.
An agent that clicks, types or enters a password in the browser now knows that a field reference needs its snapshot revision. Before, the agent often left it out, the first step failed, and the agent had to try again, so a sign-in or a form took longer.
When a Claude account reaches its 5-hour limit, an agent with automatic backup models now moves to its next ready backup. An agent with no backup waits for the reset. Before, OpenBot did not see the "You've hit your 5-hour limit" message as a limit, so each message failed with an internal error and the agent stayed on Claude until you changed the model by hand.
A scheduled, script or webhook routine run that answers only [SILENT] now stays quiet, as one that answers [[no-update]] does. Before, the chat showed the text "[SILENT]" and OpenBot sent a notification. Routines imported from Hermes Agent often use this word. A manual Test run still shows the answer.
A chat that you read no longer stays in the "Unread" group while it is open. It stays in place only while the mouse is on the sidebar or on a menu that you opened from it, so the row does not move under the pointer. It goes back to its section when the pointer leaves. On a touch screen, it goes back at once. Before, it stayed in the group until you opened a different chat.
When an MCP server refuses a tool call but the sign-in to it is still active, the agent now says so, and tells you to check the access of your account at the service. Before, the agent said that the sign-in had ended, and signing in again did not help.
An agent that answers Discord, Slack, Telegram or WhatsApp no longer stops for 100 minutes when it tries to send feedback to the OpenBot team. Before, it waited for the owner to approve a card in OpenBot. Nobody in the external chat could approve it, so that author's next messages got "Too many requests are waiting. Try again later." Now the agent gets the refusal at once and answers in the chat. On a joined server, the server must also have this version.
When you give an agent a rule for all later work in a chat, such as "always add a board card when you delegate", the agent now adds the rule to its Instructions. Before, the agent often saved the rule as a memory that it used only when relevant, so it did not see the rule in later messages and did not follow it. You can see and edit the rule in the agent's settings under Instructions. On a joined server, the server must also be on the new version.
A team no longer stops when the lead ends its turn with no answer after its teammates report back. The lead now reads the results again one time, or on its backup model when it has one. If it still gives no answer, the chat shows an error and you can send the message again.
A ChatGPT agent that ends a turn with no answer now shows an error, as other agents do. Before, the turn ended with nothing in the chat.
An agent can now save a browser screenshot as a file and attach it to its reply or to a report. Before, only the agent saw the screenshot, so an agent that tests a site for you could not keep it as evidence and stopped with a "blocked" message.
A Claude agent in a channel now ends its turn after it posts its answer. Before, Claude Code asked the agent for more text after the answer, and the agent wrote the same answer again. On a joined server, the server must also be on this version.
When the buttons of a chat card do not fit on one row, they now fill each row. Before, the main button stood alone at the end of the second row.
Grok agents can now ask a question with options. Before, the question failed and the agent asked in plain text. (Grok CLI 1.0.77)
A provider warning that does not stop the agent no longer shows "The agent could not continue. Try again." while the agent continues to work. This happened most often with ChatGPT agents on Windows.
An agent no longer stays in the "Unread" group after you read its chat. Before, unread messages in its side threads, such as the thread of a routine, kept it in the group, and each routine run put it back. Now only unread messages in the chat itself move a chat to the "Unread" group or to the top in the "Unread first" order. The row still shows a dot for unread side threads.
A GitHub Copilot agent answers again after a turn that stopped responding or that you stopped. Before, each new message failed with "Session … is already loaded" until you restarted OpenBot.
A message to a Gemini agent after an idle time now waits up to 3 minutes for Gemini to start, and longer after an earlier slow start. This also applies to an agent set to "Workspace only" or to its own account. Before, on a slow or busy computer, mostly on Windows, the message often failed after 30 s with "Gemini did not answer in time", and OpenBot did not send it again.
A Codex chat no longer sends the agent's memories and teammates again when you write after a pause of 10 minutes or more. Before, OpenBot stopped the idle Codex process, and the next message sent this unchanged context again, which used more tokens and more of the chat's context.

Nov 10, 2026

Yaji 0.67.1

Fixed

An agent that sends a message to a working teammate now knows if the message went into the teammate's running task, or why it waits. Before, the agent got "queued" every time, also when the message went into the running task, so it could not tell you what happened. When the teammate's "While working" setting is Queue, the agent can tell you to set it to Steer.
When you ask an agent to show its cursor during Computer Use, it no longer tells you that the cursor overlay is unavailable. OpenBot draws the agent cursor where the agent clicks, scrolls or moves in another app. Agents no longer get the driver's own cursor tools, which OpenBot turns off.
Agents can now use an n8n MCP server that you signed in to through OpenBot. Before, the server test showed the server as ready, but each agent chat could not connect to it, because n8n blocked the name that OpenBot sent with each request. You no longer need to add a User-Agent header to the server to make it work.
Gemini installs and updates on Windows again. Before, OpenBot 0.67.0 stopped with "The runtime archive contains a link or special file.", because the Windows download of the Antigravity CLI does not mark its program as a file.

Nov 10, 2026

Yaji 0.67.0

New

Agents now add OpenBot as a co-author to the git commits that they make, with the line Co-authored-by: OpenBot <335604065+openbotgit[bot]@users.noreply.github.com>. You stay the author. To turn it off, open Settings > Agents and turn off Credit OpenBot in commits.
The agent dock can float at the edge of the screen on Linux. It stays above other windows and on all workspaces, and clicks outside its agents go to the window under it. On an X11 desktop it works at once. On a Wayland desktop, Settings > Agent dock offers to start OpenBot again on XWayland; until then, the dock shows in the OpenBot window.
The Dynamic Island shows at the top centre of the screen on Linux. It stays above other windows and on all workspaces, and clicks outside the island go to the window under it. Hover opens it, and you can answer requests and reply from it. On a Wayland desktop, Settings > Dynamic Island offers to start OpenBot again on XWayland.
A self-hosted or hosted OpenBot server does not open the Dynamic Island or the agent dock on its virtual display. No user sees them there, and computer-use agents do not see them in screenshots.
Add a server on a Linux computer or VPS that you own with one command. Open the add server menu and select "New server on another computer". Copy the command, or copy it with instructions for an AI agent, and run it on the server. The command installs OpenBot and adds the server to your account with no questions. The app shows when the server is connected. The command works once, for 15 minutes.
Ask an agent in a chat to install OpenBot on your Linux server over SSH, for example "Install OpenBot on root@203.0.113.7". You approve the command, the agent runs it, and a card in the chat shows when the new server is connected. "Set up with an agent" in the install dialog writes this request into the open chat.
Turn on Fast mode for a Claude agent from the model picker, as for Codex, when its model supports it, such as Claude Opus 5.5. Claude gives faster answers at a higher price. Fast mode works when Claude uses an Anthropic API key, or a Claude subscription with usage credits. Before it turns on, OpenBot asks you to confirm the price: a subscription pays for fast mode from usage credits, also when the plan has usage left. When Claude answers at the standard speed, the chat tells you why. Members of a joined server do not see Fast mode for Claude, because their apps do not ask to confirm.
Server settings › Transfer can now include the chat history. Each agent's chats and side threads, with the files of their messages, and the messages of each channel move to the new server. The agent continues with the same history after the move. Turn off "Include chat history" for a small file.
A transfer now moves the channels of the exported agents, with their members, lead, icon, picture, memories, routines and archived state.
A transfer now moves workflows, the routines that a webhook starts, and the places of routines on each routine canvas. The import tells you which webhook addresses and Slack, Discord or Telegram connections to set up again.
Each agent now keeps its access, Computer Use, fast mode, backup models, hidden state, pin and notification level when it moves. A setting that gives an agent more than a new agent has, such as "Local scripts", stays off, and the import tells you to turn it on again.
Choose Ultrafast for a Codex agent. The model menu has a Speed list with Standard, Fast and Ultrafast in place of the Fast switch. Codex shows Ultrafast only for the models and the accounts that have it, such as GPT-6 Astra and GPT-6.1 Sol on a ChatGPT Pro account. On a joined server, you can still choose only Standard or Fast.
The OpenBot iPhone app can show and change the picture of a channel on this server: an uploaded image, or an icon on a color. The browser client shows the icon.
Chats with unread replies now show first. By default, they move to an "Unread" group under "Needs you", with the newest chat at the top. To change the order, open the menu of the chat list and select "Sort chats": "Manual" keeps your own order, and "Unread first in folders" puts the unread chats at the top of each folder. The chat that you read stays in its place until you open a different chat.
You can copy an image from a chat to the clipboard. Right-click the image and select Copy Image, in the chat or in the image viewer. In the viewer, you can also click the copy button beside Download, or press ⌘C (Ctrl+C on Windows and Linux).
Connect Postiz from the Apps marketplace with browser sign-in or your Postiz API key, so agents can list your social media channels, and draft, schedule and publish posts.
The OpenBot mobile app can get a notification when an agent finishes or stops with an error in a thread of this server, and a tap opens that thread. The mobile app must have its next update.
Add Honcho and Hindsight to the Apps catalog. Agents can save and recall long-term memory that other AI tools you connect can share. Sign in to your account, or use an API key.
On a joined server, the desktop app and the browser app now show the open waits of an agent above the composer, with "Cancel". The server must be on this version.
On a joined server, "Run as a new thread" on a queued message now works in the desktop app and in the browser app. The server must be on this version.
An owner or admin of a joined server now sees the programs that agents left running on the server, such as a dev server, in Server Settings > Storage, with "Stop". The command lines are redacted. The server must be on this version.
The browser app (openbot.run/app) now shows the Workflows page for a joined server. The server must be on this version.
An owner or admin of a joined server can now turn the server's providers off and on in Server Settings > Providers, in the desktop app and in the browser app. The server must be on this version.
On a joined server, the desktop app and the browser app now show, make and change routines that run once, for agents and channels. The server must be on this version.
On a joined server, the desktop app and the browser app now show the picture of a channel, and a member who sees the channel can set an icon or an image, or remove it, as in the mobile app. The server must also be on this version.
An owner or admin of a joined server can now copy the run command of a routine in the desktop app and in the browser app, when Local scripts is on for the agent. The command runs only in a shell on the server. The server must be on this version.
An owner or admin of a joined server can now turn "Use my Claude Code setup" of the server on and off in Server Settings > Providers > Claude, in the desktop app and in the browser app. The server must be on this version.
On a joined server, the queue of a side thread now shows each waiting message in full, with its actions: remove, send now and change the order. This works in the desktop app and in the browser app. The server must be on this version.
On a joined server, the usage-limit card in the chat now asks where to send a message that a usage limit holds, in the desktop app and in the browser app, for a member who can manage the agent. The server must be on this version.
An owner or admin of a joined server can now import the Hermes Agent profiles of the server's Hermes folder, in the desktop app and in the browser app. The server must be on this version.
An owner or admin of a joined server now sees the local model servers and ACP agents that the server finds, with "Add", in Server Settings > Providers, in the desktop app and in the browser app. The server must be on this version.
In the browser app (openbot.run/app), voice conversations can now use Codex live voice of a joined server. The server must be on this version. Turn on "Codex live voice" in Settings. The voice goes from the browser to OpenAI. Dictation stays in the desktop app.
An owner or admin of a joined server can now export and import the server in Server Settings > Transfer in the browser app (openbot.run/app). The server must be on this version.
In the browser app (openbot.run/app), the skills of an agent now show their marketplace descriptions and icons.
In the browser app (openbot.run/app), the search now has the "Routines" filter for a joined server. The server must be on this version.
On a joined server, the Hermes import and the local model detection now say that they read the server, not this computer.
Right-click an agent in the sidebar to mark its chat read or unread. Mark unread makes the newest reply unread, and the chat stays unread until you open it again. Mark read also reads its threads.
The pricing page on openbot.run asks you to accept the privacy policy, the terms and the right of withdrawal before you go to Checkout. The buy button stays off until you tick the box.
New setting "Learn from my browsing" in Server settings → General → Agents. It is off by default. When it is on and you control an agent's browser, the agent learns how sites work from the pages and buttons you use. Typed text, field values and password fields are not kept. The steps stay on this computer and are deleted after the agent's next memory review, or after 30 days. Turning the setting off deletes them.
An agent can tie a memory to a site, such as how to work a web page. When the agent opens a page of that site in its browser, it sees these memories with the page. The memories list in the agent settings shows the site of a memory.
The agent now keeps how a site works, such as where a button is or a step that worked after another failed, as a memory of that site. The next time it opens that site, it sees this memory and does not repeat the failed steps.
Choose the interface font and the code font in Settings > General > Appearance. Each list shows the fonts that are installed on your computer, and each font shows in its own style.
A hidden agent dock can leave a thin strip at the screen edge. Turn on Settings > Agent dock > "Show a strip at the edge", below "Hide when not in use". The strip grows as the pointer comes near, and the dock comes out when the pointer is close. Only the strip brings the dock back, so the pointer at other parts of the edge does not show the dock by accident. While an agent asks you something, a color wash goes up the strip from time to time.
Choose when Claude compacts a chat. In Server settings > Providers > Claude > Context, set "Compact chats at" to a size from 100K to 1M tokens. A lower size keeps turns fast and cheap. A higher size keeps more detail of the chat. "Claude Code default" keeps the size that Claude Code chooses for each model, as before. A model with a smaller context window compacts at its own limit. The change applies to the next Claude process. It also works for a joined server, on the web client and in the mobile app, when the server is also on this version.
The Apps catalog has 207 new apps, for example monday.com, Miro, Dropbox, Zapier, Datadog, Intercom, Klaviyo, Granola and Webflow. You connect each one with browser sign-in to your own account. You do not need an API key or a developer app. Eleven apps, such as Exa, DeepWiki, CoinGecko and AWS Knowledge, need no account.

Improved

The idle Dynamic Island stays still while the pointer is away: the greeting does not change and the logo does not blink. This uses less CPU. With the pointer on the island, both move again.
A provider change turns off Fast mode for the agent. The first start of this version turns off Fast mode one time for each Claude agent, because an earlier version did not send it to Claude. When an agent goes back from a backup model, Fast mode comes back on if it was on before.
The storage card of a hosted server now measures the storage of its plan, such as 50 GB for Standard. The system and the preinstalled tools of the server are not counted. The card shows when less than 10 GB of the plan is free, for example at 40 GB used of 50 GB. On Starter (12 GB), it shows when less than 20% of the plan is free. Before, it measured the full disk of the server and showed only when less than 5 GB of the disk was free. In the desktop app, Server settings > Storage shows the same free space.
Agents on a hosted server now know the storage of its plan. They check the space before a large download, build or container, and remove what they no longer need.
Owners and admins can now move a transfer file of up to 2 GB to and from a joined server, such as a VPS, when both computers run this version. Before, a transfer file for a joined server had to be under 100 MB. The file goes in parts, so a short connection loss does not stop it, and a new try sends only the parts that the server does not have. The chat history can also move this way. The phone app keeps the 100 MB limit.
Agents finish browser tasks with fewer tool calls: they do the steps that they already know in one browser script and do not read the page again to confirm what the script returned. On our browser tasks, this uses about 18% fewer tokens.
The agent's browser now clicks the checkbox of a Turnstile, reCAPTCHA or hCaptcha check one time before it asks you to take over the tab. This includes the Cloudflare "Just a moment..." page. Many of these checks pass with this click. An image puzzle still goes to you, and no page data goes to a CAPTCHA service.
Gemini now runs on the Antigravity CLI 1.5.4. Before, it ran on the Antigravity ACP server 1.4.5. Your Google sign-in stays: OpenBot copies it to the new CLI once.
The Gemini usage ring now reads your quota from Google. Before, OpenBot asked the agent with /usage, which the new CLI does not have.
Gemini agents show no context ring for now: the Antigravity CLI 1.5.4 does not report the context size.
The Apps tab shows the first 6 apps of each category, as the Agents and Skills tabs do. Select "See all" to see all the apps of a category. The category title and its first apps move to their new place when you open the category and when you go back.
The Marketplace shows Productivity, Documents and Design first, and Coding near the end.
Notion, Figma, Vercel, Coinbase, Obsidian, Render, Brave Search and Apple apps show their full square app icon in place of a small or round logo.
Slack, Discord, Telegram, WhatsApp, GitHub, 1Password, Bitwarden and Proton Pass show a full square icon in their brand color in the Marketplace, on their app page and in chat suggestions. Bitwarden shows its shield in place of a lock.
Agents start browser tasks sooner: the browser tools to open a page, list tabs and run a browser script are always available, so the agent no longer searches for them first.
Bug reports and ideas that an agent sends are better. Before it sends, the agent sees the open reports with a similar title, and it can add your report to one of them as a comment with what you add: your steps, your setup, and how often it happens. The report now tells the model, the reasoning effort, where the agent runs, and the recent errors of the chat. The title is in English, and the text keeps your words. The card shows the errors before you send.
An agent reports an error that it found itself only when OpenBot fails again after one retry, and the report says that the agent found it.

Fixed

On Windows, agents can use the browser while the OpenBot window is minimized or closed to the tray. Before, each snapshot image or screenshot failed with "dynamic tool request failed" or timed out after 10 seconds, until the window opened again. When a page still gives an empty image, the agent now gets the reason and can ask you to open the OpenBot window.
Tell Muse users why Muse has no models or refuses a model. Before, the model list said only that no models were available, and a failed chat showed Meta's error text. Now both say that Muse may not be available in the country where the agent runs, and that you can choose another model.
The agent browser no longer reports a security check on normal pages of sites that use AWS WAF, such as Sketchfab. Before, the agent asked you to take over again and again, and could not continue after you signed in.
When you type a name for a new or renamed folder in the sidebar and switch to another app before you press Enter, the name now stays. Before, the field closed and the folder was not created. This also let Computer Use agents create folders in OpenBot.
Start OpenBot without a burst of processes when you have many Pi agents. Before, each Pi agent started its tools, MCP servers and Computer Use at startup only to read its chat history, and kept them for 5 minutes, so the computer could stop responding.
Show the Dynamic Island on the main display of a computer that has no built-in screen, such as a desktop PC or a Mac mini, when Show on additional displays is off. Before, the island did not show on any display.
On a joined server, such as an OpenBot server on a VPS, you can now delete the history of an agent chat, and start a new conversation, compact the context and delete the history of a channel. Before, these settings showed only for agents and channels on this computer. Both the server and the app must be on this version.
On macOS, "Update & Restart" no longer quits OpenBot when it cannot install the update because the app runs from a read-only location: from its disk image, or from a temporary copy that macOS made (App Translocation). Before, OpenBot closed, did not update, and offered the same update again at the next start. Now OpenBot stays open and tells you to move it into the Applications folder.
In light mode, a switch that is on in Settings and in the skills dialog now shows a white knob on the blue track, not a black one. The blue and red buttons in Settings now show white text.
When OpenBot refuses a channel tool, for example a routine list in a task that a routine run started, the agent now gets the reason and can tell you what to do. Before, a Codex agent saw only "dynamic tool request failed".
The Codex app and ChatGPT show the agent chats of OpenBot as "OpenBot · <agent name>". Before, each of these chats had a title that started with <openbot_saved_memories>.
Workflow conditions and other workflow model steps that use Claude now work when Claude Code connects through a gateway, an API key helper, Amazon Bedrock or Google Vertex AI set in its settings. Before, these steps failed with "The model could not answer. Run the step again." while agent steps worked. Generating an agent profile with Claude had the same problem.
Work that an agent started on a backup account or model now continues on its own account or model when that limit has ended. Before, when the backup also reached its usage limit and no other backup was ready, the message failed with "You've hit your limit", the agent switched back, and you had to ask it again.
A saved two-step code from the Vault, 1Password or Bitwarden now fills the code field on sign-in pages that do not mark the field as a one-time code, such as older Keycloak pages. Before, OpenBot asked you to type the code.
The agent list in the sidebar scrolls more smoothly when you have many agents. Before, when the list reached the top or a folder name moved at the top, the app redrew every row, and the list stuttered or stopped for a moment.
Open a large group from a joined computer, the browser or the phone. Before, a group with many agents and a long history showed "The host computer's response is too large to send to this device" and did not open.
When a teammate ended its work on a request from another agent with no result and no error, the agent that asked was not told and could stay idle. Now it gets a note that no result came, so it can finish the work or ask again.
In the agent dock chat, your messages show as your own again when you are signed in, in the same order as in the main chat. Before, they showed on the agent side, and a copy of each message stayed below the newest reply.
An agent now tells you why it cannot find the models of a provider that needs a sign-in, an update or a repair, and where to fix it: Server settings > Providers. Before, the agent found no models for that provider, for example Hermes Agent without its Nous Portal sign-in, and could not say why.
The model picker now shows a new model version, such as Claude Fable 5.2, GPT-7 Sol or Grok 4.8, in the main list as soon as Claude Code, Codex or Grok lists it. The version before it, such as Claude Fable 5.1, moves under Older models. Before, the new version stayed under Older models until the public model catalog added it.
In a channel, a message that starts with "continue" or "resume" now continues a subtask that stopped while the lead waited for it, for example after OpenBot restarted. Before, the message started a new request, and the stopped subtask and its lead kept waiting.
When a Grok model gives only its reasoning and no answer, the chat now says that Grok ended the turn with no answer, and tells you to send the message again. Before, it showed the internal error empty response from model (reasoning_only). A turn that does not need an answer, such as a quiet routine run, now ends with no error.
In a voice conversation, the agent's next progress update and its answer now start to play as soon as the part before them ends. Before, each of them started only after a silence of about one second, while OpenBot prepared its voice (#6066).
On a joined server, the page of an app such as Cloudflare showed "Not connected" with no Connect button and no reason. The page now says that only an owner or admin can connect apps there, or that OpenBot on the server must be updated. While the app reads what the server allows, Connect shows as busy and then becomes available.
Show all the models of a custom agent, such as Kiro CLI, after it starts. Before, an agent that loads its models slowly showed only "Default" in the model picker until it started again.
Show the Sign in card in the chat when an agent tests an MCP server that lists its tools without a sign-in but needs one to run some of them, such as a tool that reads your account. Before, the test said that the server was ready, and those tools failed with no way to sign in from the chat.
Agents from a copied or restored OpenBot data folder now get a workspace on this computer. Before, they kept the folder of the other computer or user, and installing a skill on them failed with a permission error.
An agent can now move an agent to a model that only its second provider account serves, such as GPT-6.1-Sol on a second ChatGPT account. The model picker showed these models, but an agent that changed the settings said that the model was not available. An agent that lists the models now also sees the models of each account.
When WhatsApp does not accept a message that an agent sends, the agent now gets an error and does not tell you that the message was sent. Before, OpenBot counted the message as sent when it left this computer, also when WhatsApp refused it after that.
A deleted agent no longer comes back as an unnamed "Local teammate" agent with a message that stays queued. Before, when another agent finished work that the deleted agent asked for, OpenBot sent the result to the deleted agent and made a new agent for it.
When an agent asks again for an API key that the Vault has, OpenBot opens the key form, so you can replace a wrong key. Before, the agent got the old key at once and was told that the new key was saved.
A video or audio file in a chat of this computer plays to its end. Before, playback could stop partway, for example at 0:22 of a 37-second video, and the play button showed again.
When an agent cannot read a private GitHub repository because the OpenBot GitHub App is not installed on it, the agent now tells you to add the repository in Marketplace > Apps > GitHub > Choose repositories. Before, the agent got only "Not Found" and could not say why.
On Windows, OpenBot no longer opens a broken earlier copy with "A JavaScript error occurred in the main process" after an update. When an update installs OpenBot in an OpenBot subfolder of the default folder, it now removes the earlier copy from the default folder and points "open at sign-in" at the new copy.
An agent import keeps every skill of the export. Before, when two different skills had names that OpenBot saw as the same, such as "PPT 制作" and "PPT 审阅", only the last one stayed, and the import did not tell you. Now the second skill gets a number after its name, such as "PPT 审阅 2", and the import result tells you which skill has the new name.
When a person asks over WhatsApp, Slack or Telegram, agents no longer give the work to a channel, where the answer stayed. The agent asks the channel members directly, so the answer goes back to the person.
In a voice conversation, a noise such as a cough or a key press no longer stops the agent's reply. The reply waits while you speak, and stops only when the app hears words. Before, many replies were never read aloud.
Keep a routine run in the main chat at the place where it started. Before, when you wrote while a routine ran, the run's messages and question cards moved below the agent's newer reply when the run ended, so an old message looked like the latest one.
Sign in to an MCP server whose firewall blocks sign-in pages that name 127.0.0.1. When the browser does not come back, try the sign-in again: OpenBot then uses localhost. Before, every attempt opened the same blocked page.
A Claude agent now shows the confirmation card when an MCP app asks for your approval, such as a Composio Gmail read. Before, the request failed at once with "User declined elicitation" and no card showed.
A workflow run that starts from a step now runs a step before it again when you changed that step after its last answer. Before, a run from a step after a condition used the old answer of the condition, so it skipped the step that you chose and ended as done.
The number on an agent in the sidebar now counts only the unread messages of its chat. Before, it also counted the messages of its threads, such as a routine's thread, so it grew with each routine run and opening the chat did not clear it. Unread messages in threads now show as a badge with no number.
A Hermes agent answers again after Hermes restarts during a turn. Before, Hermes refused each later message of that chat, and the agent did not answer in its chat or in its channels.
An agent can open a browser page in a new chat again when its other chats have many open tabs. Before, the agent got "The browser can have up to 25 open tabs." but found no tabs to close, because the limit counted the tabs of all chats of the agent. Now each chat can have up to 25 open tabs, and the limit counts only the tabs that the agent can see and close in that chat.
A channel with instructions longer than 2,000 characters works again. Before, it showed "Invalid channel response", no message to it sent, and the channel list could not load. The channel keeps its full instructions, and the channel settings on the computer that has the channel show that they are too long. The web app, the phone and joined computers show the first 2,000 characters. New instructions still must fit 2,000 characters.
A chat no longer stops working after a model refuses an image because it is too small. OpenBot now starts a new session with the chat history, and the agent continues the task on your next message. Before, each new message failed again until you started a new chat. Browser images of small page elements are now large enough for these models.
When an agent searches or reads its chat history with incorrect options, it now gets the reason and corrects the search. Before, it got "History is unavailable for this request" and could not find earlier results, for example of a routine.
A webhook routine gets its endpoint also when this computer is not published as a server. Before, the endpoint did not show when you had never published this computer, or when you signed out and in again after you published it. OpenBot now adds this computer to your account by itself, and does not make it reachable online. A computer that has no server name gets the name "OpenBot".
On macOS, a Pi, Inception, DeepSeek or OpenRouter agent with "Workspace only" access can now use Apple apps and other MCP apps that start with npx or a similar command. Before, the app test passed, but the agent found none of the app's tools, because the sandbox stopped the app from writing its files outside the workspace.
Give agents the new tools of an MCP server after you test it in Server settings, or after an agent tests it. Before, an agent that was in use kept the old tool list after the server added tools at the same address, until it was idle for some minutes or OpenBot restarted.
A sign-in to an MCP server whose sign-in server answered with a web page stopped with a parser message, such as "Unexpected token '<'". OpenBot now tries the next sign-in address, as the standard asks. When no address answers in the correct form, it tells you that it cannot read the answer and that you can add an API key or a token in the server settings.
An agent's new title, name or description stays when you close its settings with Escape while you still type in the field. Before, the change was lost and no error showed.
An agent no longer asks its welcome question again in a conversation that has already started. Before, this could happen when the agent started a new session, for example after its provider stopped or after you changed the provider or the model.
The country list of a Vault contact now has all countries, for example Brazil. Before, it had only 17 countries. A country that an agent filled in now shows in the form, also when it is not in the list.
Show all your Claude models in the model picker when Claude is slow to list them at start, for example while many agents start. Before, the picker kept only the five built-in Claude models until a provider started again.
A link to a file or a folder in a folder that you added for a Workspace only agent now shows the file or the folder in the side panel. Before, it showed "Workspace file must be inside the agent workspace." An owner or admin who opens such a file from a joined computer or the phone gets the same result. Other members still open only files in the agent's workspace.
A transfer with workspace files copied the files of chat messages from the Shared folder without their messages. These files now move only with the chat history.
Moved chats and channels no longer show their old messages as unread. A moved agent shows its last message in the sidebar.
Sign in to an MCP server whose sign-in server does not let apps register by themselves, such as Gmail's, now says that the server does not accept OpenBot as an app. Before, it showed a technical message about dynamic client registration.
A ChatGPT agent now sees why a Computer Use step failed, and can tell you or recover by itself. Before, it saw only "tool_invocation_failed" with no reason, so Computer Use seemed to stop working, for example after it opened an app on Windows.
When an agent looks for WhatsApp chats to post in and the number allows no chat, the agent now knows why. It can tell you to open Marketplace, then Apps, then WhatsApp, and to add the contact or group. Before, the agent got an empty list and could not say what to do.
Agents can create routines and send feedback with a long instruction or a Windows file path. Before, the agent got "Invalid tool discovery arguments" and tried again with no result.
A Gemini chat that starts again after an idle time now waits up to 3 minutes for Gemini to start. Before, a slow or busy computer, mostly on Windows, often failed the message with "Gemini did not finish starting within 30 s".
A Grok Bot export keeps every schedule of a task. Before, the export agent could merge a task that ran three times a day into fewer runs. A task that starts on an event, such as a webhook, no longer comes over as a routine that is off and has a placeholder schedule: the export agent tells you about it and saves its prompt in imported/EVENT-TASKS.md. Paths in the instructions now point to the files in imported/. Make a new export in Grok Bot to get these changes.
A scheduled, script or webhook routine run that has nothing to report no longer shows as failed when the model sends no answer after it says there is no update. Before, Grok runs often ended this way, and you got a failure notice for a run that worked. A run that does not say "no update" and gives no answer still fails with its error.
On Windows, deleting a ChatGPT (Codex) agent works on the first try. Before, it often showed "The agent data could not be removed completely" until about a minute later or until a restart, because Codex still used the agent folder. The deleted agent's chats move to Archived in the Codex app.
A WhatsApp contact with a Brazilian mobile number gets answers again. Before, when you added the number with the 9 after the area code but the contact's WhatsApp account still used the number without it (or the reverse), OpenBot did not answer the contact's messages and showed no message. OpenBot now accepts both forms of the number. When you add a contact while the number is connected, OpenBot saves the form that WhatsApp uses, so messages from agents reach the contact.
When an agent asks a question with too many options, or with two questions that have the same id, OpenBot now tells the agent what to correct, so the agent can show the question card. Before, the agent got only "OpenBot could not create a user question." and no card showed.
Server settings > Transfer of this computer now tells you that an export is over the 2 GB or 50,000-file limit before you choose where to save it. Before, the save dialog opened, and the export then stopped with no file. The files of the Shared folder also count toward the limit.
Claude agents now connect to MCP servers that refuse requests from bots, such as some n8n servers and reverse proxies. Before, the server test passed, but the agent showed the server as failed with "Unexpected content type: null".
On a Linux Wayland desktop such as GNOME, OpenBot now shows its window again when you minimized it with its minimize button and then open OpenBot from the dock, the app list, a link or a notification. Before, the window stayed minimized and nothing seemed to happen.
A file that an agent sends to WhatsApp keeps its type. Before, a PDF, an HTML page or another document arrived as a ".bin" file that did not open. Now a PDF opens as a PDF and an HTML page opens in the browser.
The Custom tab of the model picker no longer says that the models of your endpoints load after the current task, when no task holds them. When the provider read your endpoints but lists none of their models, the tab now says so and gives an Edit endpoints button. For example, OpenCode leaves out each provider that enabled_providers in your own OpenCode config does not name.
When an agent cannot send a WhatsApp message because you did not allow the chat, or because the number is not connected, the agent now knows the cause. It can tell you to add the contact or group, or to reconnect the number, in Marketplace, then Apps, then WhatsApp. Before, the agent got only "no_answer" and could not say what to do.
A channel message that names no member now goes to the lead when the lead's routing answer cannot be read. Before, the channel often showed "The lead could not route this message. Reassign it to a member." with a "Provider error" notification, and the message waited for you to choose a member.
A ChatGPT agent now sees why a Computer Use step failed when the step gives only a text reason, such as a click on the Windows taskbar. Before, the agent saw only a general error, so it could not try a different way or tell you what went wrong.
Agents no longer save the same shared database record twice after a slow write. Before, when a write passed its time limit, OpenBot told the agent that nothing was saved, even when the record was saved, so the agent wrote it again. Now the agent is told that the write may be saved and reads the record before it tries again.
Agents on Pi, OpenRouter, DeepSeek and Inception use the skills that are installed on them. Before, these agents did not see their skills, so a skill that you added with $ did nothing, and the agent searched the computer for it.
When you revoke the OpenBot GitHub App on GitHub, GitHub in OpenBot now shows "The GitHub connection expired" and a Reconnect button. OpenBot finds the revoke when an agent's GitHub tool fails, when the GitHub page in Marketplace reads your repositories, or at the regular GitHub token check. Before, GitHub showed "Connected" for up to 8 hours, and every GitHub tool failed.
Show all the Grok models again a few seconds after Grok starts. Before, a Grok that started while it renewed its sign-in showed only Grok 4.6 and Grok 4.5 until it started again, and agents could not change to Grok 4.7.
When an agent reads the work steps of an earlier turn, such as a failed routine run, it now gets the tool steps, commands and errors that the chat shows. Before, it got "Captured work steps are unavailable" unless the agent had changed its provider after that turn.
A workflow condition with a rule in a language other than English now takes its branch. Before, the model could answer in the language of the rule, and the run stopped with "The model did not answer yes or no.". The condition now asks for an English answer, and it also accepts "true" and "false".
When Gemini lost the connection to its agent during a task, the chat showed a technical error text as the agent's answer, and the task looked complete. Your next message then also failed. Now the turn shows that Gemini stopped, and the next message that you send continues in the same chat.
OpenCode agents can create agents and use the other OpenBot tools that they find with a search. Before, the agent got "Unknown tool" and said that it could not do the task.
A Claude agent now gets the tools of an MCP app from the first turn after you sign in to the app. Before, the agent could have none of its tools for up to 15 minutes, and its session reported that the app needs a sign-in.
Run Claude workflow steps, side questions and channel summaries on the model that you selected. Before, a model from a custom Claude endpoint was changed to Claude Haiku with no message.
An agent can now create, change or delete a webhook routine when you ask for it in a thread that you opened. Before, the agent said that only a person's message can do this.
Claude agents use less of the plan limit. Before, after you or the agent edited its Instructions, a later change such as a sidebar move made Claude write the whole conversation to its prompt cache again. Now an Instructions edit goes with the next message, and the agent follows it from that message on.
OpenCode connects after its download when its first start fails one time. Before, OpenBot showed "OpenBot could not update the OpenCode CLI. The new version did not start.", removed the new copy, and then showed that OpenCode is not downloaded. Now OpenBot keeps the copy and starts OpenCode again after a few seconds.
Shared tables stay available after a write that ran too long. Before, the stopped write could make the table list empty, and each read or new table failed with "attempt to write a readonly database" until the agent changed a row.
You can now import a Grok Bot export of more than 100 MB into a joined server, such as an OpenBot server on a VPS. The app sends the export in parts, so its size is limited only by the free disk space of the server. Before, the import refused the file with "the export must be a .zip under 100 MB". Both the server and the app must be on this version.
When OpenBot refuses a message from an agent to a teammate, for example a direct message from a channel task or an unknown recipient, the agent now gets the reason and can correct the call. Before, a ChatGPT agent saw only "dynamic tool request failed" and tried the same call again.
A Claude agent that you or a teammate stops now stops within about 10 seconds, and its queued messages start. Before, when Claude did not end the stopped turn, the agent showed as working for hours and the messages behind it never started until you restarted OpenBot.
A channel no longer goes quiet when the usage limit of a member's model stops its work. The channel now shows the same card as the agent's chat: it names the model at its limit and lets you move the member to another model, or wait for the reset. A message that you send in the channel also lets the member try one turn, as in its own chat, so work starts again at once when the limit has ended.
The usage limit card now names the model at its limit, such as "Fable 5.2 usage limit reached", and the usage ring and panel show that model as used up with its reset time. Before, the card named only the provider while the ring and panel showed usage left, because a limit of one model is not in the provider's 5-hour and weekly windows.
A message that could not be sent now shows the reference of the failure beside "Not sent", so support can find the cause.
The dock's + menu no longer shows a bottom fade when all its agents fit.
When you asked an agent to change the reasoning effort of an agent on a model with no effort setting, such as some OpenCode models, the agent said that only "Medium" was supported, and it showed a saved effort that had no effect. Now the agent tells you that this model has no effort setting, and it does not show an effort for that model.
A browser script that returns a list named elements now gives the agent its result. Before, the script ran its steps but the agent got only the error "element.states is not iterable", and it could run the same steps again.
On very large web pages, an agent can now find and use the buttons and fields that are on the screen, such as an upload field. Before, the agent's page read stopped early and left out visible controls, and a CSS selector could fail with "CSS selector uniqueness scan exceeded the safe node limit".
Agents now find saved memories when they search for a word in Chinese, Japanese, Korean, Thai or another language that does not put spaces between words. Before, most of these words found nothing, so the agent said it did not know a fact that it had saved.
A change between the light and dark themes no longer moves a wash across the agent dock and the browser dock. Before, each dock showed its own wash, out of line with the main window. The docks now change at once.
On a Linux server, the agents' browser now opens sites that need WebGL, such as the Figma editor. Before, these sites stopped with a WebGL error.
In the full browser view, the message box no longer covers the middle of the web page. Before, a button or link under the message box, such as "Try another method" on a sign-in page, did not get the click. The message box is now at the bottom right (bottom left in right-to-left languages). Click the arrow in the message box to shrink it into a small tab, and click the tab to show the box again. You can do this at any time, also while the agent works. The tab shows what the agent does, and a dot shows a reply. When the agent asks a question or asks for a code, the message box comes back.
On macOS, the agent dock stays on screen when you hide OpenBot (⌘H). Before, it hid with the main window.
GPT live voice now speaks the agent's reply when the reply is long. Before, the voice got only a part of a long reply, and its answer could be different from the reply. A long reply is now spoken as a short summary of its facts, and the chat shows it in full. This applies to voice conversations on the phone and on the computer.
When you type @ in the dock chat, the list of agents now shows above the message box, not behind the chat.
In the browser app (openbot.run/app), a joined server, such as an OpenBot server on a VPS, now shows the context meter with "Compact now" for agent chats and side threads, and "Delete history" for agent chats. Channels show "New conversation", "Compact context" and "Delete history". Before, only the desktop app showed these settings. The server must be on this version.
In the desktop app, a joined server now has these features. Before, only this computer had them. The server must be on this version: - The "Files" filter of the search. - Custom ACP agents in Settings and in the agent form, for an owner or admin. - "Change key" for the server key, for the owner. - The place where the routines of an agent send their results, for an owner or admin. - Import of passwords from a CSV file into the server vault, for an owner or admin. - The memory limit in the agent settings, for an owner or admin. - "View" and "Undo" on an instructions change in the chat. - The provider account in the new agent form, for an owner or admin. - Messaging places, Telegram chats and Slack workspaces as workflow triggers and outputs, for an owner or admin.
In the browser app (openbot.run/app), a joined server now has these features. Before, only the desktop app had them. The server must be on this version: - The "Files" filter of the search, and search results in side threads. - Custom ACP agents in Server Settings > Providers, for an owner or admin. - "Change key" for the server key, for the owner. - Import of passwords from a CSV file into the server vault, for an owner or admin. - The provider account in the new agent form, for an owner or admin. - Memories and the Routines row in the agent settings. - Folders, the instruction file, the routine results place and the memory limit in the agent settings, for an owner or admin. - The Files pane of an agent and of a chat. - "View" and "Undo" on an instructions change in the chat. - "Mark as unread", "Send now" on a queued message, "Send to channel", MCP sign-in from a chat card and the Usage link in the chat. - "Leave server" and "Remove server" in Server Settings > General. - The disk warning card in the sidebar, for an owner or admin.
On a joined server, the message search now also finds messages in side threads. The server must be on this version.
On a joined server, an owner or admin can now send a workflow output to an OpenBot channel. The server must be on this version.
In the browser app (openbot.run/app), the sidebar now shows the unread count of an agent after "Mark as unread", and the unread dot of the server shows only when a chat has unread messages.
On Windows, when OpenBot stops an idle Cursor agent, or another agent that starts from a .cmd file, it now stops all of the agent's processes and MCP servers. Before, some of them could keep running in the background and use memory (#6470).
A chat no longer stops working after a model refuses an image because it is too large. OpenBot now starts a new session with the chat history, and the agent continues the task on your next message. Before, each new message to that agent failed with "Inline image exceeds maximum dimensions" until you started a new chat.
An agent that gets an update from a teammate no longer writes filler in your chat. When the update has nothing for you, the agent's text stays under the "Update from" marker, closed, with no notification.
When a saved login could not finish and you then took over or cancelled, the browser tab stayed on "Secure authentication". The agent could not read or close it. Now the page loads again when you give it back, and the agent can use the tab.
A Hermes Agent answer no longer shows two times in one message. Before, when a Hermes plugin changed the answer, the chat showed the first answer and then the changed answer after it. On a joined server, the server must also have this version.
An agent no longer asks you to connect GitHub when GitHub is already connected. It uses its GitHub tools, or gh and git, instead.
Keep template placeholders such as {{name}} and links such as [[Page]] in a Pi agent's reply. Before, the chat and each quote of the reply showed [redacted-unscanned] in their place.
On Linux, Arabic text in the app and on openbot.run now uses Noto Sans Arabic when it is installed. Before, the system often drew Arabic in a font that was hard to read.
Agents find their saved memories when they ask for more than 10 results. Before, the search failed and told the agent to change its search words, so the agent could not recall the memory.
An OpenCode subagent that works in the background no longer shows "Working" with a timer that never stops after OpenCode stops reporting on it. Before, the card stayed so until OpenBot started again. Now OpenBot keeps the chat's OpenCode session open while the subagent still reports work, and the card shows "Stopped" when the session closes.
The GitLab app shows its logo again. The GitLab site refused the icon request.
When a provider such as Gemini did not start, its diagnostics showed only "Internal error", and the reason that the provider sent was lost. Now the diagnostics and the log show that reason. When the start was not part of an update, the provider row shows it too.
A routine that sends work to a teammate now continues when the teammate answers. Before, the agent read the answer without the routine's task, so it only reported the answer and did not do the next step, such as sending the work to the next teammate, until you wrote to it.
Browser scripts fail less often. An agent can choose one of several matching elements by its position, such as the second "Delete" button. Scripts also accept more ways to wait for a page, such as waiting until a loading spinner goes away, can resolve the links that they read with URL, and can read the address of the tab with location.
The Gemini model list now shows the models in Google's order, best first, such as Gemini 4 Argon first on a Google AI Ultra plan. Before, GPT-OSS 120B came first, and new Gemini agents ran it.
When a provider stops offering an agent's model, the agent now moves to the newest model of the same family, such as Claude Opus 4.6 to Claude Opus 5.5. Before, it moved to the first model in the list. Between two models of one version, an agent on a thinking model takes the thinking model.
A model such as GPT-OSS 120B no longer comes first in a model list: its size, 120B, is not a version.

Nov 9, 2026

Yaji 0.66.0

New

The "What's new" notes after an update show in your app language: Arabic, Chinese, Dutch, French, German, Hebrew, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish and Turkish. Notes of releases before 0.56.0 stay in English.
A workflow output can post its result into an OpenBot channel. Choose "Channel" in "Send to": the channel list shows the OpenBot channels first, then the channels of each Slack, Discord, Telegram or WhatsApp connection. In an OpenBot channel, the result shows as a note of the workflow. It starts no task, and no agent answers it.
The agent menu of the Tasks composer has a search field. Type part of a name to filter the agents, and press Enter to take the first match.
On the Tasks page, drag the edge of the task panel to make it wider or narrower, for example to split the window between the list and the chat. The app keeps the width. Double-click the edge to go back to the default width.
On the Tasks page, the message box of an open task has a microphone. Speak a message to the task, as in the chat.
Help → Report a problem opens the bug report form on GitHub with your OpenBot version and system already filled in.
Agents can send a bug report or a feature idea for OpenBot for you. When something in OpenBot goes wrong, or you say what you want, your agent offers a short card with the text in plain words. Select Send to the OpenBot team: OpenBot opens the issue on GitHub, adds the app version and your system, and a thank-you card in the chat links the issue. You do not need a GitHub account, and the issue does not show your name. When you are not signed in, you get the filled-in form to submit yourself.
Agents can post Discord embeds: a card with a title, a description, a link, a color, fields, a footer, a time, an author, a thumbnail and an image. Ask the agent to post as an embed. Plain-text posts do not change.
Click an image that you attached to a message to see it at full size before you send it. This works in chats, in threads and on the Tasks page.
Vault logins can keep a two-step (authenticator) code. Open a login, select Set up, and paste the setup key that the site shows under its QR code, or choose a screenshot of the QR code. OpenBot shows the code to enter on the site before you save, then shows the live code next to the login in the list; select it to copy it. A CSV import also reads the two-step keys of a password manager export, and adds them to logins that you imported before.
Agents can sign in with a saved two-step code: they enter it with the secure sign-in card, and they never see the key. When you ask an agent to turn on two-step verification on a site, it saves the site's setup key on that login in the Vault.
Owners and admins of a joined server see and set up two-step codes of the host's Vault when both computers run this version. If you go back to an older OpenBot version, its first change to the Vault removes the saved two-step keys. The logins and their passwords stay.
Admins can edit a custom endpoint on a joined server: its name, address, models, key and headers. You no longer remove the endpoint and add it again, so agents stay on their models. The server must run this version of OpenBot.
Add channels to the agent dock. Use + on the dock to add a channel. Click a channel to read its newest messages and send a message to it. The channel shows its unread count, its work, and when it waits for you. "Open in the main window" opens the channel there.
Agents now learn skills by themselves. After a correction or a method that worked, the background memory review saves what the agent learned as a new skill, or adds it to one of the agent's own skills. The chat shows "Created skill" or "Revised skill".
Server settings > General > Agents > Skill learning: choose Automatic (the default), Ask first, or Off. With Ask first, the agent offers to save the skill; click Save on the offer, or answer yes. Save writes the skill in the background, so it costs a small part of a full agent reply. The offer and Save also work on a joined server and in the web client.
With Automatic, an agent now fixes its own skill while it works: when you correct how it follows the skill, or a step fails and it finds the fix, it saves a new revision and starts to use it. The chat shows "Revised skill". Ask the agent to undo the change, and it goes back to the revision before.
Read replies aloud with Eleven v4 or Eleven v4 Turbo, the ElevenLabs models that speak more than 90 languages. Open Settings > Voice, and select the model in Spoken replies.
Select many files in the Library and act on them at once. Drag a box over the files, click Select, or Command-click (Control-click on Windows and Linux) a file. Shift-click picks a range. A bar at the bottom downloads the selected files to the Downloads folder or deletes them, and shows the progress. Members of a server can select and download files.
The + menu of the agent dock has a search field when 9 or more agents and channels are not on the dock. Type part of a name to filter them: the list changes size smoothly as the rows change. Press Enter to add the first match. Only the rows scroll, and Dock settings stays at the bottom of the menu.
Agents can watch online videos, such as Facebook, Instagram, TikTok or YouTube links. They get the title, the author, the counts, still frames and, when speech recognition is set up in Settings → Voice, a transcript of what is said. The video downloads to a temporary folder on your computer and is deleted after the agent reads it. The sound goes only to the speech recognition model that you chose. Only public videos work.
On Windows, the Dynamic Island now shows agent status and items that need you at the top of each display, as on a Mac. Turn it off or change its size in Settings > Dynamic Island.
On Windows, the agent dock now shows your agents at the edge of the screen, above the other apps, as on a Mac. Choose its place and its agents in Settings > Agent dock.

Improved

A new agent now opens with its own short welcome, in your language, as a question with three first tasks that you can choose or answer in your own words. The welcome stays in the chat after you answer. The setup message that OpenBot sends to the agent no longer shows in the chat.
The agent's working line, such as "Waiting for your answer", now shows under an open question or approval card, not above it.
While a new agent prepares its welcome, the chat shows that it is working, not an empty "ready" page. When you skip the welcome question, the agent replies with one short, friendly line.
In a channel, a greeting, a question about the channel or an unclear message goes to the lead, which can ask what you mean. Before, the Coordinator could answer it itself and stop the task.
Channel members do the work they get themselves. They give work to another member only when you ask for that member or for help, or when a part clearly needs another member's role.
A channel that you opened recently shows its messages at once when you open it again, without Loading.
A stopped task in a channel shows its owner and the start of the request, so you can tell several stopped tasks apart.
When channel members pass a task between them too many times, the card names them and says that the task stopped to prevent a loop. Before, it said "The automatic assignment limit was reached", in English only.
With the Natural response style, a channel member's progress note, such as "I'll check the shared folder first", and its answer show as separate messages, as in an agent chat. Each message shows once, when it is complete. Before, the text grew in the channel while the member wrote it.
Channel members no longer start a message with their own name, such as "Research: ..." or "Launch check: ...". The chat already shows the name above each message.
In a channel, the activity line folds away when the work stops. Before, it went at once and the messages above jumped.
A new message in a channel slides in, as in an agent chat. Before, channel messages appeared in one frame. With the Natural response style, the message enters with all of its text. Its words do not fade in one by one, so a complete answer does not look as if it still arrives.
When the lead gives parts of one request to several channel members, the desktop app shows one row, such as "Assigned to Chief, Builder and Launch". Before, it showed one row for each member.
A channel member answers in the channel while it works in its own chat, as in a side thread. Before, the channel waited until the chat turn ended, and the chat waited for channel work. An agent still does one channel task at a time. Agents whose provider runs one turn at a time keep the old order.
When one channel member stops while others work, its face shrinks out of the activity line and the other faces close the gap. The line keeps its label until all work stops.
The theme ring has more light effects. Each color moves a different distance, so colors split at the wave. The page bends at the window edges when the wave gets to them. Borders and dividers light just before the wave crest. The theme sweep also splits colors. The ring uses the same frame time as before.
The Settings page is now called "Your settings", so it is clear that its settings are yours and not the settings of the server. A row at the end of its section list opens the settings of the server in the top row, and the server settings page has a row that opens your settings again. This is the same on the desktop app and in the web client.
The server settings page has a search field, on the desktop app and in the web client. It shows only the sections whose name or description contains the text. Enter opens the first section.
Connect Composio with a sign-in in the browser. You no longer type an MCP URL and an API key. When an agent needs an app, such as Gmail or Slack, it asks you to connect that app. To use a key, select "Use an API key instead" in the connect dialog. The key is the only field.
When you find the models of a custom provider, OpenBot now also sets the efforts that the endpoint lists for each model, for OpenRouter and RefuseLess. The effort menu then shows only the efforts that the model accepts, for example Max, and no Medium when the endpoint refuses it.
When an agent's Slack post pings a person, that person can reply in the thread without mentioning OpenBot. The reply goes to the Slack Orchestrator, with the post and the name of the agent that sent it. Before, a reply with no mention reached no agent. In that thread, replies of other people always need a mention.
In a Slack thread, an agent can stay quiet when a message needs no answer, such as thanks. OpenBot then removes its "Working on it…" post and keeps the ✅ reaction.
The CPU and memory use in the Usage header is live: it updates every few seconds while the Usage page is open. Before, it updated each minute.
The Roles page of server settings is now named Permissions.
A browser or phone connection to this computer that goes quiet for a few minutes, such as when the phone app is in the background, now continues when it comes back. Before, the connection ended after about 2.5 minutes, and the device made a new one.
The web client connects to a server faster, because it starts the next steps while it waits for the account service.
The macOS app icon has the same look as the mobile app icon: a lavender tile with a soft gradient and glass eyes. On macOS 26, the Dock shows the glass, dark, clear and tinted versions of the icon while OpenBot runs too. Before, the Dock showed a flat image after OpenBot started.
A Grok Bot or Hermes Agent export that you import on this computer has no size limit, and it can have up to 200,000 files. Before, the limit was 2 GB and 50,000 files. OpenBot now reads the file from disk one file at a time, so a large export does not use memory, and it reads ZIP64 files, which zip tools make for files over 4 GB. An agent whose workspace files need more free disk space than the computer has is skipped, with a message that names both sizes. A damaged file in the export is skipped with a warning, and the agent still imports. An import into a joined server still accepts a file of 100 MB or less.
An agent import or a transfer file that has .git or node_modules folders now leaves those folders out, with one warning that counts their files. Before, OpenBot refused the whole file. The Grok Bot export agent also leaves out other folders that a tool makes again, such as .venv, dist, and build.
OpenBot now keeps up to 100 MCP servers. Before, the limit was 32. To manage more than 32 MCP servers of a joined server, update OpenBot on the device that you connect from.
When you correct how an agent does a task, it applies the correction at once and does not stop to save a memory. The background review keeps the lesson, so the reply comes sooner and costs fewer tokens.
Open the export agent in Server settings → Import → Grok Bot now opens the new Migrator by OpenBot page in Grok Bot.
The cards in a chat that wait for you have one design: the approval, browser takeover, browser code or password, usage limit, task answer, Vault request, Marketplace suggestion and MCP sign-in cards. Most of them show "Waiting for you" while they are open. Skip or Cancel replaces the X, and all buttons are pills.
In a narrow chat, such as a pop-out chat window, the status of a card goes under its title and the buttons fill the footer.
Agents no longer show rows of reply chips. They ask with a list of options instead. Reply chips in older chats show as a list of options, and you can still answer them.
While an agent works, the Dynamic Island and the moving agent avatars use less CPU and GPU.

Fixed

Delete history of an agent chat works again. Before, it stopped with "Invalid SQLite column rowid." when the chat had messages, and kept them.
A change of only the chat color or the exact color of an agent on a remote server no longer shows "Could not save agent settings". Before, the color was saved, but the app showed this error.
In the desktop app on macOS, times now follow the "24-hour time" setting in System Settings > General > Date & Time. Before, chat times, routine times and usage limit notifications showed "2:30 PM" in an English interface also when the Mac used a 24-hour clock.
Agents can now use a remote MCP server that you signed in to, such as n8n, when it uses the newest MCP version. Before, the agent showed "failed to connect" and had no tools from the server, but the server test passed.
Codex no longer shows "The agent could not continue" when you turn on an under-development feature, such as async_compaction, in the Codex config.toml. This Codex warning does not stop the agent. OpenBot now writes it to the log only.
When agent settings or an avatar cannot be saved, the error now shows only where you made the change: in the settings panel, or in the message box for a model or fast mode change. It gives the reason from the server. If you close the panel before the save ends, a notification shows the error. Before, the same error also showed as a message in the chat.
A menu in Settings that shows its current value again no longer makes the window stop responding. Before, a menu such as "Requests an hour" in Settings › Agents › Local scripts could save its own value again and again after you changed another setting, and you had to restart OpenBot.
The Download page on openbot.run opens again. Before, the Download button in the top bar showed "This page could not load", and a direct link to the page showed an empty page.
When a Muse or ChatGPT sign-in fails, the provider row now shows the reason from the provider, such as a Meta account that has no payment method. Before, the row showed only "OpenBot could not connect Muse. Try again."
In a narrow sidebar, the icon that shows that an agent waits for you is now a round badge. Before, it was a small, uneven pill.
The start animation now fills the window when the window changes size, for example when it goes to full screen. Before, it showed only a small part of the crew in a corner of the window.
The start animation now plays one time when the account or a server answers slowly. Before, the crew fell a second time when the workspace started to load.
On the New agent page, the mouse wheel now moves the row of suggested agents. Before, a mouse wheel did not move the row, so you could not get to the suggestions on the right.
Kimi agents can now revise skills and run tools that need numbers or lists, such as data queries with parameters. Before, Kimi sent these values as text, and the tool refused the call.
Show the models of your Mistral account in the model picker. Before, Mistral offered only Mistral Medium 3.5, unless you added models to ~/.vibe/config.toml and restarted Mistral.
The ElevenLabs app can connect. Before, ElevenLabs showed "Authorization Error" in the browser, and OpenBot showed "Waiting for the browser…" until the sign-in stopped.
When an agent waits for an answer to a question card and you write a new message in the same chat, the agent now reads your message. The card closes as skipped, and the agent can ask again. Before, the agent did not reply until you answered or skipped the card. Approval cards and browser sign-ins still wait for your answer.
When a network proxy stops a Gemini (Antigravity) reply with an HTTP error, such as "HTTP/1.0 502 Bad Gateway", the chat now says that the service did not complete the request and that you can try again. A proxy's 429 error now shows the rate limit message. Before, the chat showed only the raw error text.
A question that a new agent asks in its first turn now shows at once, not only after a reload.
Restart after an update now waits until the ring from the button has crossed the window, so the whole animation shows before OpenBot quits.
An update that downloads automatically now shows the same ring as a click on Download.
Agents on Cursor and on some other providers now keep their OpenBot tools after OpenBot restarts. Before, a chat that OpenBot read again after the restart could lose them: the agent could not make routines, save memories or message teammates, and no error showed.
A channel message that starts with a quote and then an @mention now goes to the mentioned member. Before, the mention was ignored and the lead chose who answers.
A channel message that starts with several typed names, such as "@Ana @Ben", now goes to each of them. Before, only the first name counted.
A channel member that you name in a message, such as "thanks @Ana", now answers. Before, routing could decide that the message needed no answer.
Channel members no longer get the narration of a turn that already has an answer. This makes channel turns use fewer tokens.
Work in another channel no longer makes the open channel read its messages again on each update.
The "New" line in a channel no longer moves to each answer that comes while you read the channel. It stays where it was when you opened the channel, as in an agent chat.
A message that you send in a channel shows once. Before, it could show twice for a moment after you pressed Enter.
A Codex agent now gets the tools of an MCP app or connector that has the same name as a server in your own Codex settings. Before, a tool list in those settings could hide all of the app's tools, while its test said that it was ready.
A sent message no longer shows an extra @ before a pasted file or an MCP server tag. A file whose name has a [, such as [Draft] plan.pdf, now shows as one chip, not as broken text.
Channel list previews, awaiting replies, search results and the message queue show a pasted file by its name, not as @image.png or raw markup.
A message that one agent sends to another in a side thread no longer makes the thread unread. Before, each of these messages added 1 to the unread count of the thread.
When an agent sent a teammate several requests and they joined one running turn, only the request that the teammate answered got a result. The other requests now get the result of that turn too, so the agent that asked does not wait forever.
OpenBot no longer stops with "Render frame was disposed before WebFrameMain could be accessed" when a page loads while another browser tab has a crashed page.
A ChatGPT agent finds the skills in its own workspace. Before, it sometimes looked for a workspace skill in your personal skills folder, then said that the skill was not available.
An agent can show an email address again when it reads it from its memories or from earlier messages. Before, it saw [redacted-email] and could not give you the address. Secrets stay hidden, and logs still hide email addresses.
A ChatGPT (Codex) chat no longer compacts again and again after the agent saves a memory or you change its settings. Before, the next message could wait in the queue while the agent showed that it was working, and each extra compaction used your plan.
With the Natural response style, a step that runs longer than one minute, such as a long command, now shows how long it has run on the activity line. Before, the line did not change, and a working agent looked like an agent that had stopped.
The backup models list in agent settings shows a backup that had a provider error a short time ago as "Not available now · skipped". Before, the list showed it as ready, but the switch to a backup model passed over it with no message.
A custom ACP agent that ends a turn with no answer now shows an error in the chat, not an empty chat (#5110).
Agents make fewer failed browser calls. The browser tools now take more of the argument forms that agents write from other browser libraries, such as a ref or a CSS selector as a plain string, wait_for states such as networkidle, and tab_id. A browser script can wait with a function that reads the page, and it can read a whole page in one evaluate step.
When you open an agent's chat while the agent is still writing its reply, the chat now follows the reply to the end. Before, a chat with new messages stayed at the "NEW" line and the reply ran off the bottom of the page until you scrolled down.
A message that you send while a routine or a webhook runs in the chat now says that a routine is running. Before, the queue said only that steering did not succeed.
A routine run now tells the agent to do all of the task in one run. Before, an agent could stop after the first step with a progress report, and you had to send a message for each step.
Search shows the agents whose name matches first. Before, an agent that had the word only in its description or last message could show above it, and the agent you named could be hidden in All.
Show a teammate's answer in a side thread as an agent message, as the agent's own chat shows it. Before, the answer showed as a full chat bubble with all its text.
The + menu of the agent dock now scrolls when you have more agents than the dock window can show. Before, the agents at the end of the list and Dock settings were cut off, and you could not add those agents.
You can add a shared agent again after you delete it. Before, the install stopped when one of your local skills had the same name but different text. Now the agent's text becomes a new version of that skill, and your other agents keep the version that they use.
When OpenAI cannot start a Codex live voice call, OpenBot now says that the problem is on OpenAI's side and that you can try again later. Before, it showed a technical "404 Not Found" error with a web address.
Show the total of a Cursor plan first, as Cursor shows it, with the "Cursor models" and "Other models" lines under it. The Cursor ring shows this total, or a line that is used up. Before, the ring showed only the two lines, so its numbers did not agree with Cursor (#5848).
An agent's queue no longer starts the same message again and again while the agent waits on a question that it asked itself. Before, a routine turn that ended in the same chat made the chat look idle. The queue then tried to start a turn several times a second until the question ended. One server wrote 66 GB to its database in a day this way, and the disk became full.
A chat database no longer grows when a queue only moves. A change of queue position is shown at once, but it is not saved again for each queued message. A chat update that the next update replaces whole is no longer kept in the history.
Retry on a chat or a workspace panel of a remote server now checks the connection and connects again when the server does not answer. Before, the server showed as online, and each Retry failed again after 15 seconds until you restarted OpenBot.
A web client connection that restarted its network path many times no longer drops at the next restart.
When you close the "Usage limit reached" notice for a ChatGPT plan, it stays closed in that chat until the limit resets. Before, it came back after a few seconds while the agent worked on credits.
Show no "Mistral error" message when Mistral Vibe sends a request again after a dropped connection. Before, each retry line in its log showed as an error, also when the turn then completed. A turn that fails still shows its error in the chat.
Keep the Claude usage ring on the left rail when Claude cannot read the plan usage for a short time, such as after too many reads. Before, the ring went away until a later read worked, and OpenBot did not wait longer before it read again.
The page of a role in Server settings › Roles now shows the access that its members really get. Before, an agent showed "No access" even when Everyone gave all members "Can manage", so a role seemed to hide agents that it did not hide. When Everyone gives more, the agent shows that level with "· Everyone", and lower levels are off. "Change in Everyone" in the menu opens Everyone.
Extra provider accounts, such as more Gemini accounts, no longer ask you to sign in again when the computer is busy. Before, an account check that took more than 30 seconds showed "Sign in again" for each account, also when its sign-in was still good.
A ChatGPT agent no longer stays on "Compacting context" with no end when the connection stops during a compaction. After 20 minutes OpenBot stops the turn. When the connection fails or stops during a compaction, the agent continues the task once by itself. When it fails again, a plain message tells you to check your connection and send a message to continue.
An agent can now create or revise a skill when its description has a colon, such as "Use when: ...". Before, the save failed and the agent had to try again.
A skill can now have a name with no Latin letter, such as a name in Japanese. Before, the save failed.
Show the monthly Kimi Code quota in Usage and the usage rail. Before, a Kimi plan with a spent monthly quota showed 100% left, because only the 5-hour and weekly quotas were read.
When you ask an agent to use a Chrome window that you already use, and OpenBot cannot attach to that browser profile, the agent now tells you this in plain words. It then continues the task in the OpenBot browser, where you can sign in. Before, the agent showed driver error codes and stopped.
The Usage panel now says "left" after each percent, such as "100% left". Before, it showed only "100%", so a limit with all of its usage left looked like a used limit. A long limit name now ends in "…", so the bars stay visible.
A channel request that names a member as typed text, such as "please check this, @Ana", now goes to that member. Before, the lead could end the request with no answer and no message (#5964).

Nov 8, 2026

Yaji 0.65.0

New

Read replies aloud with ElevenLabs (Flash v2.5, Multilingual v2 or Eleven v3) or xAI, with your own API key. Choose the model in the Spoken replies section of Settings → Voice. Supertonic stays the default and runs on your computer. Listen and read-aloud replies use the chosen model.
On a Mac, Settings › General › App behavior has "Hold ⌘Q to quit": While agents work (the default), Always, or Never. With Never, one ⌘Q quits, also when agents work (#5500).
Open a page that an agent makes as a side panel tab, next to the chat. Point to the page in the chat and click the side panel button.
Select text in an agent's page, then click Quote or press the quote shortcut to quote it in your message. The quote opens the page in the side panel.
Choose a conversation background color for the Light theme and another one for the Dark theme, in Settings › General › Chat background. The setting changes the color of the theme that shows now. When the theme changes, the conversations use the color of that theme. A color that makes text hard to read becomes darker or lighter. Default keeps the canvas of the theme.
Agent settings and channel settings have a "Delete history" action. It deletes the messages and files of the chat or channel from OpenBot, and the agent or the channel members forget them. The setup, side threads, memories, routines, and tasks stay.
Add Mistral as a provider. OpenBot downloads Mistral Vibe 3.1.0 and runs its ACP server. Sign in with a Mistral account, or set MISTRAL_API_KEY. OpenBot turns off Vibe's telemetry, and Workspace only agents do not change your Vibe settings in ~/.vibe.
Groq dictation: select Groq Whisper large-v3 turbo or Groq Whisper large-v3 in Settings → Voice and save your own Groq API key. OpenBot sends each recording to Groq when you stop, with the language that you set. The key is kept encrypted, in a file of its own.
Codex live voice now works with agents on a joined or hosted server. The server's Codex sign-in starts the call, and the audio goes between your computer and OpenAI. Before, the voice of these agents was only transcribed. The server must run this version of OpenBot.
A Library page in the window rail shows every picture and file from the chats of a server, newest first. Pictures and files show as a grid or as a list. You can search them, filter them by type, open a picture or a file in a viewer on the page, and go to the message that has the file.
A queued message to an agent that cannot steer, such as Gemini or OpenCode, now has a Send now button. It stops the agent and sends the message at once. Tab in the message box does the same.
Pi, Inception, DeepSeek, OpenRouter and custom ACP agents can search the web with the openbot.web_search tool. A separate Codex session on the Codex sign-in of this computer runs the search and returns a short answer with source URLs. Agents on other providers keep the web search of their own CLI.

Improved

Each section of Settings → Voice has one Model row with a grouped menu and one Download or Remove button. A cloud model shows a key row below it. Download progress shows in a toast with Cancel.
Show a model or agent step after a workflow join which branch has no answer, and why: a step failed or stopped, with its error. Before, the join ran on the other answers, and the step could not see that an answer was missing. Approvals, conditions and outputs still get only the answers.
The Anthropic API key settings now say that Anthropic bills the key's API credit, not your Claude plan.
In the model menu, the Anthropic API key account has a key badge in place of a number, so you can see it apart from your Claude sign-in.
Notifications have a new look. The main action is a filled button at the bottom right, with the secondary action as text before it. The close button is a small round button on the top right corner, and it shows when you move the pointer over the notification. The GitHub star, social post, delete undo and slow server start notifications use the same look.
When you send a message, the chat moves to your message at once, and the message slides in. Before, the chat glided down and a long message pushed the chat.
The message box grows from one line when your text wraps, and stops at 6 lines. Then it scrolls. When you send or clear the text, it goes back to one line in one smooth move.
The working agent shows about 0.75 s after you send, so it no longer comes in while your message still moves.
In the light theme, the selected agent, channel or person in the sidebar has a grey fill, so you can see at a glance who you talk to. Before, its white card was hard to tell apart from the sidebar.
Agent browser tabs find bot checks, such as a Cloudflare "Just a moment..." page. When the check passes by itself, the agent waits for it, for up to 15 seconds. When the check needs a person, the agent asks you to take control or reports it, and does not try to solve it or load the page again.
An agent browser snapshot waits up to 2 seconds while the page shows loading placeholders. Before, the agent read the placeholders and had to look at the page again.
An agent that opens a browser tab gets the page with it, and the first browser result brings the other browser tools. A short browser task now often takes 1 or 2 steps instead of 3 or 4.
Agent browser scripts accept more wait forms that models write, such as sleep(500) or waitFor({ target, state: "visible" }). Before, these scripts failed and the agent tried again.
OpenCode, Antigravity, Cursor, Cline, Kimi, Devin, Copilot, Hermes, custom ACP agents, Muse, Pi and DeepSeek agents get their saved memories and section-mates with the messages, as Claude agents do. A memory edit or a sidebar move now sends one small update, not the full instructions again, so the provider's prompt cache stays valid.
Claude agents use fewer tokens in each request: about 4,000 fewer (24%) without your Claude Code setup, and about 2,000 fewer (12%) with it. Claude Code skills that cannot work in OpenBot, such as schedule, loop and send-to-project, are off. Coding skills such as code-review and claude-api stay available, but agents see only their names. Remote Control tools are off, and so are the Claude Code tools and helper agents for the Claude Code app, such as its review panel and status line setup.
Commits and pull requests that Claude agents make no longer add a Claude co-author line. Your own attribution and skillOverrides settings, in your Claude Code setup or in the project's .claude settings, still apply, and your own skills keep their listing.
Pi, Inception, DeepSeek and OpenRouter agents use fewer tokens in each request: about 800 fewer (9%) with Mercury 3. The section about Pi's own documentation is no longer in the system prompt. The tools that ask for an API key and suggest a Marketplace app now load through tool search, as other OpenBot tools do. Your own Pi settings, extensions and skills still apply.
OpenCode agents send about 12% fewer prompt tokens in each request. OpenBot turns off the OpenCode question tool, as agents ask you with their own question tool, the two MCP resource tools, and the report and opencode skills. A rule for these in your own OpenCode config, or in the project's opencode.json, still wins.
Grok agents send smaller requests. OpenBot turns off the Grok turn summary, an extra request after each turn that OpenBot does not show, and Grok no longer loads the skills in ~/.claude/skills. With one developer's skills, each request is about 2,800 tokens smaller.
OpenBot now keeps the Grok settings in your ~/.grok/config.toml. A value that you set there for feedback, workflows, the turn summary or Claude skills applies to OpenBot agents too.
Hermes agents use fewer tokens in each request: about 2,400 fewer (16%). The Hermes browser tools are off, because each agent has its own OpenBot browser. In a code workspace, agents see only the names of Hermes skills that are not about code, and can still use them. Your own coding_context and disabled_toolsets settings in the Hermes config.yaml still apply.
The "Steer agents while they work" setting no longer says that it works only with ChatGPT and Claude. More providers steer now; agents that cannot steer keep new messages in the queue.
A voice conversation reads the agent's progress updates aloud while it works, each one when it is complete, and then the answer.
Claude and Codex agents send fewer tokens with each request. OpenBot tool descriptions and teammate messages are shorter, and tool schemas no longer show validation-only limits to the model. A Codex request is about 800 tokens smaller, and the OpenBot tools of a Claude request are about 1,900 tokens smaller. Tools and their behavior do not change.
Channel members get the saved channel memories again only when the memories change. Before, each turn sent all the channel memories again.
A Claude agent keeps its saved memories and section list in the conversation after an MCP server, skill or setting change. Before, the next turn sent all of them again.
After a restart, the background memory review reads only the messages after its last review. Before, the first review after a restart read up to 200 messages again.
Memory reviews, channel summaries and profile drafts now run at the lowest reasoning effort that the model takes, on each provider whose model has effort levels. The effort of the agent's own turns does not change.
On Claude, model calls without tools no longer send the Claude Code coding instructions. This includes memory reviews, channel summaries, profile drafts, side questions, workflow model steps and channel routing. Each call sends about 2,300 fewer input tokens, and spends less of the Claude plan.
Codex agents get their saved memories and their section teammates with the messages, not in the system prompt. When a memory or a section changes, Codex keeps its session and its cached prompt, and gets only a short note about the change. Before, a section change started a new Codex session, and a memory change reached the agent only in its next new session.
Agents now get browser screenshots at the page size, not at the screen resolution. On a Retina screen, each screenshot uses about one third of the tokens. The browser panel still shows the full resolution.
Large evaluate and run_script values and large GitHub tool results now reach the agent cut to a fixed size, with a note that tells the agent how to get the rest. Before, each large result stayed in the conversation and made each later turn cost more.
When an evaluate or run_script value is larger than 64 KB, the error now tells the agent how to get the value in smaller parts. Before, the agent often stopped at this error.
Codex agents use about 1,050 fewer tokens (about 5%) in each request. Codex no longer offers to install ChatGPT plugins, which OpenBot cannot complete, and no longer lists the openai-docs, work-pets and plugin-management skills, which are about the Codex app. A tool_suggest or [[skills.config]] setting in your own Codex config.toml still applies.
Titles, profiles and other short background texts that Codex writes use about 1,500 fewer tokens.
Use less memory for agent browser tabs. Only the 3 most recently used idle tabs stay loaded; an older idle tab unloads its page after 2 minutes and loads it again on its next use. An idle tab now unloads after 15 minutes, not 30. Before, each open tab kept about 300 MB until it was idle for 30 minutes.
A browser preview card no longer loads an unloaded tab again when the tab has no saved preview.
Keep fewer chat images in memory: the image cache is now 16 MB of text, not 64 MB.
Daily resource analytics now show the JavaScript heap, page object heap, ArrayBuffer memory and DOM node count of the app windows, and on Windows the private memory of each process group.
Idle Claude, ACP (such as OpenCode) and Pi threads now stop their processes after 5 minutes, not 10, and only one idle thread keeps its process. Its next message resumes the session.
On a server that is low on memory, a Claude thread whose background tasks sent nothing for 30 minutes now closes.
Codex agents now free the memory of idle conversations: OpenBot unloads a Codex session 5 minutes after its last turn, and keeps only the newest idle session of each Codex process loaded. It does not unload a session that runs a background terminal or waits for an approval or an answer. The next message resumes the same conversation.
The app window keeps the messages of only the 5 most recently opened hidden side threads in memory. An older side thread loads its messages again when you open it.
A browser preview card asks for a new picture every 3 seconds while the agent works, and every 20 seconds when the agent is idle.
Streamed reply text now updates the chat once per frame, not once per received part.
The Markdown cache of chat messages now keeps at most about 2 million characters, so long pasted logs do not stay in memory after you close their chat.
Played voice replies keep their audio for 3 minutes. A later replay makes the speech again.
A live browser view that is not on screen stops receiving frames.
When OpenBot stops Claude Code, Codex or Pi, it now also stops the processes that their MCP servers started. Before, a child of an MCP server could continue to run, and use memory, after Claude Code stopped.
Each app window now loads only English and the language it shows, not all 15 languages. This uses about 30 MB less JavaScript memory in the main window and in each helper window.
Claude Code errors show the end of the CLI's error output again when the CLI stops with an error.

Fixed

Claude agents on an Anthropic API key now use the key's promotional API credit. Before, Anthropic counted their requests as Claude Code, which promotional credit can leave out, so a key with only such credit failed with "Credit balance is too low".
When a Claude agent uses an Anthropic API key that has no API credit, OpenBot now says that the key has no credit, that a Claude plan does not pay for API key requests, and how to use the Claude plan. Before, it showed only "Credit balance is too low".
In a channel, Quote now puts the selected text in the message box.
The queue above the message box closes in one smooth move in the chat, in a side thread, and in the browser panel. Before, it stopped for a moment and then jumped.
A send no longer stops the chat for a moment the first time a turn plays the classic completion sound.
A message that you send to an agent that does not work no longer shows in the queue for a moment before it goes into the chat.
In a channel, a file link from an agent now opens a file that the agent used outside its workspace, as in the agent's own chat. Before, the link showed the error "Nothing exists at … in the agent workspace …".
Workspace only Claude agents start again when ~/.claude/settings.json has env, apiKeyHelper or another sign-in setting. Before, each message failed with "Cannot use both a settings file path and the sandbox option".
When the page in an agent browser tab crashes, the next browser step loads the page again, and a running browser script stops at once. Before, the tab stayed broken, and the script waited until its step timed out.
Sign in to an MCP server that refuses 127.0.0.1 redirect addresses at registration. OpenBot now registers again with localhost. Before, the sign-in stopped with "Client registration failed." before the browser opened.
On macOS, agents can use the browser while the OpenBot window is minimized or closed to the menu bar. Before, each snapshot image or screenshot of a page that loaded in that time waited 10 seconds and failed, until the window opened again. Hide OpenBot (⌘H) still stops these images.
The agent dock now scrolls when it has more agents than the screen can show, and its top and bottom edges fade when more agents are behind them. Before, the agents below the screen edge were cut off. An agent that you open or add scrolls into view.
On macOS, window switchers such as AltTab no longer show the agent dock and the Dynamic Island as OpenBot windows.
A desktop AppImage on a Linux computer that also runs a self-hosted server updates itself. Before, each download failed with download_failed because the app wrote its request to the server's private folder.
A server now connects to Signal again when Signal does not answer its socket. Before, the server could stay "online" while devices got "The host is offline." until a restart (#3749). The server waits 30 seconds for each answer, and checks a connection with no traffic for 4 minutes. After 3 minutes with no connection, the server status and openbot status say that other devices cannot reach it. The message goes away when the connection comes back.
Gemini agents no longer fail with an empty-output error when they end a turn without text after a tool call or on a message that needs no reply.
Start a turn on the Pi provider when the developer instructions are long. Before, the turn failed with Pi process closed (1) on Windows, because the launch line passed the 8191-character limit of cmd.exe. The instructions now go to Pi in a file.
The Docker image now ships Noto CJK fonts, so Chinese, Japanese and Korean text renders in the embedded browser. Before, those characters showed as hollow boxes.
Remote desktop on a self-hosted Linux server or in the Docker image shows a white desktop before an agent opens a window. Before, it showed a black screen, as if no video came.
A duplicate of an agent keeps the named browser profile of its source. Before, the duplicate used the shared browser session.
A routine run can now read routines with openbot.list_routines, so a task such as routine-healthcheck can inspect agents' routines. Before, the routine wrapper prompt told the agent not to list routines, and the task could not read them.
Usage now shows the cost and the input token counts of a custom agent. Before, a custom agent showed no cost, and its input and cache counts were blank when the agent left out a cache count of 0. Thanks to @Muzhen2000 for the root cause.
Change a setting on the Local scripts page of Settings › Agents without a stop of the window. Before, a change such as the Local scripts default made the window stop responding, and the app used a full CPU core until it was closed (#5614).
A server that you connect to over HTTP or HTTPS now connects again when the network comes back after a Wi-Fi sign-in page. Before, a sign-in page that answered in place of the server stopped all retries, and the server stayed disconnected until you selected Retry.
A server that was online in the last 10 minutes now connects again within a minute after it comes back. Before, when its computer or network restarted, the app waited up to 5 minutes (15 minutes without focus) before it tried again.
Routines run again when the computer wakes but OpenBot does not get the wake event. Before, they stayed on hold until OpenBot restarted.
When Google asks you to verify your account, a Gemini turn now fails with a clear message, and the error banner has a Verify account button that opens Google's page. Before, the turn showed Google's text as an answer. (#2419)
A failed tool step in the steps of a turn now has a "Show error" button that shows the error of the tool. In the Trace, a click on a failed tool step shows its error under the step. Before, the step showed only "failed", and the agent seemed to try again for no reason.
In the compact sidebar, the top and bottom edges of the agent list now fade clearly when more agents are above or below. Before, a cut avatar looked like a hard edge.
In a channel with the Natural response style, each agent's progress updates now stay in the chat as that agent's messages, as in the agent's own chat. Before, an update showed and then went away into the agent's steps.
In a channel, a Pi, DeepSeek, OpenRouter or Inception agent's note before a step no longer shows as a message and then goes away. The channel shows the note when it is complete.
On a Retina screen, a click at a point that an agent read from a browser screenshot now hits that point. Before, the click went to a point at twice the distance.
An agent that ends a short answer with its next work, such as "Let me run the tests." or "I'm going to fix the parser.", now continues by itself. Before, only a few phrases, such as "I'll now …", started the automatic continuation. An agent that asks "Would you like me to continue with the remaining files?" now also continues, also when its plan has open steps.
The Dynamic Island and the agent dock no longer animate a working agent while you use other apps. Their loops made macOS draw the screen again on each frame, so other apps could stutter while an agent worked. Motion starts again when you move the pointer over the island or the dock, or open a chat from the dock.
The + menu of the message box now has one Attach files item. Before, it showed Attach image and Add context, which both opened the same picker for all file types.
A file that an agent names as Shared/… in a message now opens from the shared folder when you click it, and its image shows in the message. Before, the app looked for the file in the agent's own folder and showed that the file was not found. (#5657)
A saved login now fills on sign-in pages whose password field hides the characters with page styles in place of a standard password field, such as some bank sign-in pages. Before, the password card opened each time and you typed the password by hand.
A browser tab that you open now stays open when its page does not load, so you can see the address and try again. Before, the new tab closed. When a site has a security certificate that is not trusted, OpenBot now says so in plain words. (#5630)
When a Cursor plan is used up, an agent set to switch models without asking now moves to its backup model. Before, Cursor's "Upgrade your plan to continue" showed as the agent's answer, and the backup model did not start. Other agents now show that the Cursor plan is used up, in place of that answer.
The agent dock above other apps now shows an agent's reply while the OpenBot window is closed or minimized. Before, the dock showed only your message until you opened the OpenBot window again.
Agents now use a login that you save in the Vault, 1Password, Bitwarden or Proton Pass after their chat started. Before, an agent that started its chat with no saved logins did not know about them, opened the sign-in page and asked you to type the password.
Before an agent asks you to take over a sign-in page, it now checks your saved logins first.

Nov 7, 2026

Yaji 0.64.0

Action needed after you upgrade

Settings > Providers > Claude has a "Use my Claude Code setup" switch. It is off by default: the agents on this computer no longer load your own CLAUDE.md, skills, agents, hooks and plugins from ~/.claude, so each turn sends a shorter prompt. The files of an agent's workspace still load. Your sign-in settings in ~/.claude/settings.json, such as env and apiKeyHelper, still apply. Turn the switch on to give the agents your Claude Code setup again. If your Claude agents use your own CLAUDE.md, skills or hooks, turn on Use my Claude Code setup after you upgrade.

New

A server owner or admin can now log in a server's Codex with an OpenAI API key from Settings > Providers > Codex > API key. Before, the key option showed only for Codex on this computer. The server must run this version.
Choose who can control this computer with remote desktop: the owner, owners and admins, or every member. Find it in Server settings › Remote desktop. Every member can still control it until you change the setting.
Turn on Ask on this computer to allow or deny each remote desktop session at the computer. A session that nobody allows in 30 seconds does not start.
While someone controls this computer, OpenBot shows a notification and a menu bar or tray item with their name. Stop all remote control in that menu or in Settings ends every session.
The Usage page now shows a cost for Pi, DeepSeek, OpenRouter and Inception turns. The cost is the one that Pi calculates from its model list, not what a subscription charges. Before, these turns showed tokens but no cost.
Choose what each custom provider model can read beside text: Images, and PDF on OpenCode. A Pi model with Images selected now gets the images that you attach and the images that the agent reads. Before, a Pi custom model got text only, and the agent answered that the model does not support images. An OpenCode model with PDF selected can read PDF files. You can also clear Images for an OpenCode model whose endpoint refuses images.
Find models now fills the context window, the max output and the accepted inputs of each custom provider model when the endpoint lists them, as OpenRouter, vLLM and gateways in the models.dev format do. A value that you typed or saved stays. A line under the list tells you which values come from the endpoint.
Connect GitLab.com from the Apps marketplace with browser sign-in, so agents can work with projects, issues, merge requests, repositories and CI pipelines. For a self-hosted GitLab, add its MCP server address in the MCP servers settings.
A thread tab now shows Load older messages when the thread has more than its newest 100 messages. Before, the older messages of a long thread could not be shown. On a joined server, this needs a host on this version.
Read the openbot.run home page in Polish. Open openbot.run/pl, or set Polish as the first language of your browser. The header, the menu and the footer of the home page are in Polish too.
Read the Teams, Pricing and Contact pages of openbot.run in Polish, at openbot.run/pl/teams, openbot.run/pl/pricing and openbot.run/pl/contact. Links on a Polish page go to the Polish pages.
Choose English or Polish with the language button in the openbot.run header, on the home, Teams, Pricing and Contact pages. The website keeps your choice for one year.
The openbot.run footer links to OpenBot on Threads, LinkedIn, Reddit, Instagram, TikTok and Facebook.
The openbot.run home page shows two more features: multiplayer mode, where people and agents work in shared channels, and workflows, which join agents, models and approval steps.
Connect Omnisend from the Apps marketplace with browser sign-in, so agents can read campaign reports and work with contacts, segments, campaigns and automations. Select read-only access on the Omnisend consent screen when agents only need reports.
Connect ElevenLabs from the Apps marketplace with browser sign-in to turn text into speech, work with voices, and manage ElevenLabs voice agents. Each generation uses ElevenLabs credits.
Sign in to MCP servers that accept only a client metadata document and have no client registration, such as ElevenLabs.
Add an agent dock on macOS. It is on after you upgrade, and it stays at the edge of the screen, above your other apps. Put your agents on it from the + button, then click an agent to chat with it in a panel that grows out of the dock. Each agent on the dock shows the same states as in the sidebar: working, waiting for you, unread answers and failed runs. Right-click an agent to remove it from the dock, or to open it in the main window or in a new window. Drag the handle below the dock up or down to move the dock along the edge; each display keeps its own position. Settings → Agent dock turns the dock on or off, chooses the agents on it, the display it shows on (or all displays), the left or right edge, and whether it hides in the edge until you move the pointer there. It can also stay in the OpenBot window instead of above other apps.
Click the microphone in a dock chat to dictate a message. Click send while it records to send the words at once.
When an agent on the dock finishes an answer or a file, a short card shows beside the dock. Up to three cards show, and "+N more" shows the others. Each card hides after some seconds, and comes back while the pointer is on the dock. Click a card to open that chat. Turn it off in Settings → Agent dock → Show finished work.
The chat now shows what an agent says between its steps, such as "I found the config, now I check the logs", as messages while the agent works. Its thinking stays in the steps. This works for Claude, Codex, Grok, Pi and the ACP agents. Grok's thinking also stays in the steps. Agents now write a short update only when a finding changes the plan or before a long step: what they learned and what is next. This is the new "Natural" response style of a server. To get the old behavior, choose "Classic" in Server settings → Personality.
With the Natural style, an agent that uses a tool first writes one line: what it understood and what it checks first. Its final answer starts with the verdict, explains technical terms in a few words, and says what it could not check, and whether a read-only task still wrote files.
While a Claude agent runs a command or a subagent, the activity line shows the agent's own short label of that step, such as "Find three largest folders in ~/Library/Caches", in place of a general text.
The activity line of Codex, Grok and the other agents now names the running step from its command, such as "Reading the file main.ts…", "Measuring folder sizes…" or "Visiting github.com…".
While a tool step runs, the activity line keeps its label and no longer changes to "Waiting for…" after a few quiet minutes.
With the Natural style, a Claude agent gets the opening-line rule again with each message, so a long chat keeps the short first line.
With the Natural style, the "Working on it…" post in Slack, Discord, Telegram and WhatsApp shows the agent's newest progress line, such as "Measuring folder sizes first", until the answer replaces it. Secrets are removed, and the agent's thinking never goes there. The Stop button stays.
With the Natural style, the activity line of an agent chat shows only what the agent does now, such as "Checking Poland's weather". It has no general label, timer or step list. Each new step label rises into place, and the label stays while the agent reads the step's result. The avatar shows what the step does: a curious look for a search, wide eyes on a web page, a focused look while it reads, an eager one while a command runs, and the thinking dots between steps. The Classic style keeps the old line.
Use OpenBot in Hebrew. The layout mirrors from right to left when Hebrew is the language.
Members of a joined server can now manage its MCP servers and connect apps such as Trello, not only owners and admins. A member can change, sign in to, or remove only the servers that they added. Each app still has one sign-in that all agents of the server use. Both the server and your app must have this version.
An agent can now tell what it did in its other chats. When you ask "what did we do today?", it lists its own chat and its side threads, with the last request and reply of each, and reads the one it needs. When you ask it to find something that you said in another chat, it searches all of its chats in one step. A side thread can also read the agent's own chat. Before, an agent saw only the chat that you asked in.
The website has a Brazilian Portuguese copy of the home, teams, pricing and contact pages. Select "Português (Brasil)" in the language menu, or open /pt.
Channel settings have "New conversation" and "Compact context" for a channel on this computer. A new conversation makes the members forget the earlier messages, and their next turns send a much shorter prompt. The messages stay visible above a divider, and the members, instructions, memories and routines do not change. Compact now makes each member keep a short summary of its channel session. It works for members whose provider OpenBot can ask to compact, such as Codex and Pi. Claude members keep their own automatic compaction.
When the message box is empty, Send in an agent chat now starts a voice conversation. Click it again to end the conversation. When you type, Send shows again. + > Voice also stays.

Improved

A server owner or admin can now open a file that an agent wrote outside its workspace, such as in ~/Desktop, from the phone or the web client. This works when the agent has no workspace limit, as on the desktop. Members can still open only files in the workspace.
Agents now write the full path of a file outside their workspace, so a click or tap on the file name opens the file. Before, a reply could name only test.md, and the file did not open.
Agent browser tabs refuse cookie banners automatically. They click only a reject button, never accept. Tabs that you open keep their banners, and the refusal stops when you take control of a tab.
New Grok agents now start on grok-4.7, and the model picker marks it as the Grok default. Agents that exist keep their model. A Grok CLI that does not list grok-4.7 starts new agents on the first model in its list.
The rail at the side of a long chat now shows one short tick per message you wrote, at equal spacing. With a mouse, the ticks show when the pointer comes near the rail. Point at a tick to see a card with the message and the start of its reply; the ticks beside it grow a little. Click the tick to go to that message.
Ping a Slack user that an OpenBot post mentions as <@user ID>, for example a question from an agent to a teammate. Before, the post showed the mention as text and sent no notification. Channels, @here, @everyone and user groups still get no ping.
New text on the openbot.run home page: the first section, the agent section and the feature section.
The openbot.run header shows the menu button up to a window width of 1100px, not 900px, so that the links do not overlap the logo in Polish.
A theme change now moves across the window in a soft wave of color, instead of a sudden switch.
When you click Download or Restart on the update button in the sidebar, a ring of color moves in to the button or out from it. Restart waits half a second so that you can see the ring. Nothing moves when the window is hidden or when your system asks for reduced motion.
A message longer than 20,000 characters shows its start and a "Show full message" button. Before, the app drew the whole message, and a very long agent message could stop the window for many seconds.
You can now attach a file of any type to a chat, for example a VPN profile (.ovpn), a certificate (.pem) or a Terraform file (.tf). Before, OpenBot refused a file whose extension was not in its list. The agent gets the file's path and reads it with its own tools. Open shows the file in the file manager. A remote server must run this version of OpenBot to take these files.
A Grok Bot or Hermes Agent export can now be up to 2 GB and 50,000 files, the same as a transfer between OpenBot computers. Before, the limit was 500 MB and 5,000 files, so the export agent refused larger workspaces. An import into a joined server still accepts a file of 100 MB or less.
A hosted server that was asleep connects sooner after you open it. The server starts OpenBot sooner, publishes itself to the account service right after it signs in, and the app connects as soon as the server is ready instead of at its next check.
After a wake, a hosted server checks its providers with a longer time limit while its disk is restored, and looks for provider updates after the restore. Before, a provider that worked could show that the computer was busy.
A hosted server no longer saves its npm download cache when it goes to sleep, so a wake restores less data.
A new hosted server no longer installs Hermes and Cursor by itself when OpenBot starts. To use them, select Download on their provider row. Before, each server saved about 3 GB of these CLIs, and each wake restored them.
The installed app uses about 150 MB less disk space: it no longer includes the source maps of its dependencies.
After a server starts, the phone, the web and other computers now show its agents as soon as the agents are loaded. Before, they waited until each AI provider had started, which took about 40 seconds on a hosted server after a wake.
The app download is smaller: it no longer holds copies of libraries that are already built into the app, such as the diagram and icon libraries. A hosted server also has less data to restore when it wakes.
When product analytics is on, the startup summary now also records how long each startup step took, rounded. Before, it recorded only the slowest step. It records no names, paths or content.
Ping a Telegram or WhatsApp user that an agent's answer or post mentions: @username or @<user ID> in Telegram, @<user ID> (the phone number) in WhatsApp. Before, Telegram showed the mention as text, and WhatsApp sent no notification. An answer with a mention is a new message, because an edit of "Working on it…" pings nobody.
The "Run in a separate thread" switch now shows in the routine settings of an agent on a joined server, on the desktop app and in the web client. Any person who can manage the agent can use it. The server must run this version or later. Before, the switch showed only for agents on this computer. The iPhone app does not show the switch yet.
On a Mac, ⌘Q asks you to hold it or press it twice only while an agent works. The hint now names the agents that the quit stops, with their faces. With no agent at work, one press quits OpenBot, as in other apps.
In an agent chat, the steps of a turn now show under the answer, as a channel shows them. Select them to open the trace. With "Show agent steps" on, the open list stays above the answer.
A message now shows only React, Reply and a More button (⋯) when you hover it. More opens the same actions as a right-click: Copy, Read aloud, Send to channel, Branch, Mark unread and Delete.
An agent chat no longer shows the time under each message. A centered time shows above a message on a new day or after a pause of one hour or more.
Messages sent one after another, by an agent or by you, now read as one block: where two bubbles meet, their corners are smaller. All the updates of one agent turn stay in one block, also when a long step keeps the agent quiet for more than 5 minutes.
The actions of your own text messages now show beside the bubble, on its left, so your messages sent one after another stay close together. A message with images keeps its actions under it.
To get the full action row and the time under each message again, turn off Settings → General → Compact message actions.
A running shell command that showed "Working through the next tool-assisted step…" on the activity line now shows "Running a command…". Before, a command that does not return looked the same as any other step.
The agent list of the skill dialog shows the title of each agent next to its name, as the sidebar does.
A sign-in that a joined server sends to this computer, such as for a Marketplace app, now opens the sign-in page in your browser by itself. Before, the page opened only when you clicked "Open sign-in page". The button stays, so you can open the page again. The web client still opens the page only when you click.
An agent in a messaging chat, such as WhatsApp or Slack, can no longer read the owner's other chats. Other people can write in a messaging chat.
A change of language in the website menu no longer loads the page again. The page changes to the new language at once.
On macOS, the agent dock now shows the first 3 agents of the sidebar until you add or remove an agent. Before, the dock was empty until you added agents one at a time.

Fixed

A removed server member stays removed when another admin changes their role at the same time. Before, the role change could make the member active again.
A voice recording that started after the microphone stopped the previous one by itself now runs to the full time limit. Before, it could stop early.
Stop now ends an agent turn after a short wait even when the agent does not answer the stop request. Before, the turn could stay open.
On Windows, OpenBot works normally after another app cancels a sign-out or shutdown. Before, closing the window then quit OpenBot instead of keeping it in the background.
When you select other text while an instruction for a selection is being sent, the new selection keeps its action bar. Before, the result of the first send could close the bar or show an error on it.
An attachment download from a server that the server stops, or that arrives damaged, now fails at once. Before, it waited one minute for a timeout.
A hosted server that you wake no longer goes back to "stopped" for a short time. Before, a status check that started before the wake could replace the "waking" state and send the start request again.
A connection to a server now fails at once when the server disconnects during the connection handshake. Before, the app waited 30 seconds before it showed the error or tried again. Also, the app no longer starts new server connections while it quits.
A hosted server that you join again no longer shows "Starting" for a long time. Before, a wake reply that arrived while the server was being removed left a start that never ended.
An agent that you open in a new window now uses the full width of the window. Before, when the servers showed as a rail in the main window, the new window had an empty strip at the left.
A Claude Code agent now recovers when Claude Code cannot find the saved session of its chat. The message that finds the problem fails once with an explanation, and OpenBot starts a new session with the chat history, so the next message works. Before, each later message failed with "No conversation found with session ID", and Continue failed the same way. This occurred, for example, when the first message after a move to another Claude account failed at the terms prompt, or after a change of the Claude config folder.
Antigravity turns of an agent that an earlier version saved with an effort in its model name, such as gemini-3.8-flash-high, now show a cost. Before, they had no cost.
A channel with long agent thoughts opens again from a phone, the web or another computer. Before, the server refused to send it with the error http/413/response_too_large. The first page now holds fewer messages, and you load older ones as you scroll up.
The prompt cache time on the Claude page of Settings › Providers now updates when another admin or the mobile app changes it. Before, the page showed the old time until the app started again.
In Settings → Providers, a model switch in the Models section no longer makes the section flicker or the page jump. Before, each change hid the model list for a moment and then showed it again.
A server that you removed from your account no longer comes back by itself. Before, the computer that runs it published it again within a minute, and the server returned to the server list of each member. That computer now stays private and shows that the server was removed. Publish it again on that computer to share it.
A selected server that cannot connect no longer keeps the whole window on the "Connecting to OpenBot" screen. After the first failed attempt, the server shows its connection screen beside the server list, so you can select another server.
Ping a Discord user that an OpenBot post mentions, for example a reminder. Before, the mention showed the user but sent no notification. Roles and @everyone still get no ping.
Stop chats from showing through each other and the window from flickering when Transparent window is on, on macOS 26 and later. OpenBot now builds the window with the blur when the setting is on. If you turn the setting on while OpenBot runs, restart OpenBot once.
"Keep memories in this channel" now works on a joined server and in the browser client. Before, the switch always showed off there, and turning it on saved nothing. A server that is not updated hides the switch.
Chats open, scroll and update faster, most of all on slower computers. Before, each new message, row or reply text in a chat made the app restyle almost the whole window, because some styles watched the whole page for open panels.
A turn that ends without an answer, for example one that you stop or that a restart interrupts, now shows its steps in a row of its own in the agent chat, in channels and in thread tabs. Before, its thinking was not shown after the turn ended, and a chat or a thread whose newest messages were only that thinking showed no rows.
Keep the database small when a channel member streams a long message. Before, each update of the message wrote a full copy of it to the event history, so a long thought added gigabytes. Messages from before this update keep their earlier copies.
When a provider's quota runs out after an agent starts work on a message, for example a spent Gemini quota, the agent now continues that work on its first ready backup model. Before, the message failed with the usage error and the agent stayed on the spent model. This also applies when the agent is set to ask before it switches, and when the first backup's quota is spent too.
Sign in to MCP servers, such as ActiveCampaign, that let OpenBot ask only for the scope it registered. Before, the authorization page showed invalid_scope for offline_access, and the sign-in did not finish. OpenBot now registers offline_access with the other scopes. When the server reports an earlier registration without it, OpenBot registers again.
A Pi, DeepSeek, Inception or OpenRouter turn no longer keeps running after its final answer when you saved a short password or API key. Before, the turn stayed active and the next queued message waited until the app restarted.
A chat opens with less work when its draft has no tags. Before, the message box read its scroll position each time a list of the chat loaded (skills, MCP servers, tabs and threads), and each read laid out the window again.
After you add OpenBot to Slack or Discord, the "Open OpenBot" button on the openbot.run return page now shows also when a browser extension or a script error stops the rest of the page. Before, the page showed only "Return to OpenBot", and the connection did not finish.
The desktop log now records why a Slack or Discord connection from the browser did not finish.
A chat or channel with one message larger than about 1.5 MB opens on a phone, in the web app and on a joined computer. Before, the chat showed "response too large" and the older messages did not load. The device now gets the start and the end of that message, up to 256 KB, with a note that tells how much is not shown. The full message stays on the computer that runs the server.
A new Claude chat uses a little less of the Claude usage window. Claude no longer sends an extra request to name each new session, because OpenBot now gives the name.
When a provider, such as Grok, cannot use an MCP server from your MCP settings, OpenBot now shows an "MCP server not started" message that names the server. A server that asks for a sign-in tells you to sign in to it, and the message has an "Open MCP settings" button. Before, the provider's raw text showed as a provider error, and the provider row showed "Error" while the provider worked. A server whose name was not in that text showed no message at all. Each server shows its message once while OpenBot runs.
In the desktop app, the agent name or avatar in a teammate message now opens the side thread that sent the message, or else the agent's own chat. Before, it opened the side thread that the agent showed last, which was often an unrelated thread.
An agent that answers a teammate's request with an unknown message ID now gets an error and can send the answer again. Before, the answer went to the teammate's main chat, not to the side thread that asked, and that side thread waited for an answer that never came.
On Windows, a message with an attachment, such as a pasted image, no longer fails with "OpenBot does not have permission to complete this action" when antivirus or the search indexer opens the new file for a moment. OpenBot now tries again for about 3 seconds before it shows the error.
A hosted server that went to sleep after a wake now closes OpenBot cleanly. Before, OpenBot stopped with its database still open, and the next start checked the whole database first, which made it slower.
Show the approval and question cards of two Codex agents on different Codex accounts at the same time, and send each answer only to its own agent. Before, the second card replaced the first, the first agent waited forever, and an answer to the first card ran the request of the other agent.
Let an agent fill a saved Proton Pass login in the browser. Before, submit_secret refused the Proton Pass login id from list_logins, because the id was longer than 128 characters.
A hosted server now starts after a stop that left a lock file of its virtual display. Before, each start of OpenBot failed after 20 to 40 seconds, again and again. The stop now waits until the display closes, and a start removes a lock file that no process holds.
You can now add an agent or a skill from the Marketplace to a server that you made on this computer. Before, Add stopped with http/409/team_api_error, because that server has no account session. The desktop app now reads the listing with its own account.
Proton Pass logins now work after OpenBot restarts. Before, each read failed with "Could not connect Proton Pass. Connect again." until you disconnected and connected Proton Pass again.
A Proton Pass login with no password, for example one with only a passkey, now gives "has no password for this site" when an agent uses it to sign in. Before, OpenBot reported that the password manager could not be read, and signed in to Proton Pass again.
On Windows, a Gemini (Antigravity) chat now ends its helper process when it is idle. Before, each Gemini chat that you opened since the start of OpenBot kept a helper process of about 120 MB, so OpenBot used more memory and became slow after hours of use. The first message after a chat was idle can now wait while Gemini starts.
An agent chat that you open again, after a channel or another page, stays where it opens. Before, it scrolled back to the message that you opened earlier from search, a notification or a reply.
A chat that a compaction left at 70% to 79% of the context window compacted again after each turn, without end and without a notice. One chat compacted 25 times in 27 minutes. Now, after 2 such compactions in a row, OpenBot stops automatic compactions of that chat and shows the notice. "Compact now" still works.
The web app now opens, saves and sends files of up to 100 MB, the same limit as the desktop app. Before, a file larger than 10 MB did not download.
When the web app cannot download a file, it now shows the reason at once. Before, it waited one minute and then showed "The attachment download timed out."
A file larger than 100 MB now shows "The file is larger than 100 MB." in place of a general error.
A teammate's question or answer now comes back to the Slack thread, Telegram reply chain, or WhatsApp chat that asked, also when the teammate sends it without the id of the request. Before, it went to the agent's own chat, and the person in the conversation did not see it.
A stop word such as cancel or stop, or the Stop button, now stops a Slack, Discord, Telegram or WhatsApp request that is still starting. Before, the request ran to the end and posted its answer, and the stop word started a new request.
When an agent tries to stop a teammate's work and a task of the teammate still runs, the agent now learns that the task continues and that you can stop it from the task tab. Before, the agent was told that nothing ran.
The Custom tab of the model picker now shows the models of each custom endpoint that runs on Pi in their own section, with the endpoint name. Before, the models of all Pi endpoints were in one list with no heading, and two models with the same name could not be told apart.
OpenBot no longer closes without a message when an agent's browser action waits on a page that navigates, for example a store page with payment frames. Before, the app could crash and stay closed until somebody opened it again.
Sites that check for bots no longer find the built-in browser as easily. Before, a tab in the background or in a minimized window reported a window smaller than its page, window.chrome was empty and frames had none, and a tab that was not shown kept a fixed 1200x800 page size.
Agents now type short text one key at a time with real key codes, and move the pointer along a short path and hold the button when they click. Before, text arrived with no key events and the pointer jumped to the target with no hold. Clicks in a tab that is not shown take no longer than before. Text with a line break, a bracket or a quote is still entered in one piece, so code editors do not add closing brackets to it.
On a phone, the website language menu went past the left edge of the screen. It now stays on the screen.
The agent pages of the Marketplace on openbot.run open again. Before, a click on an agent showed "Not in the Marketplace", because the site accepted only agent ids in one form, and the agents that OpenBot curates use a different form. Links to these agents in the app work again too.

Nov 6, 2026

Yaji 0.63.0

New

Connect Higgsfield from the Apps marketplace with browser sign-in to generate images, videos and audio, and check credits and assets. Each generation uses Higgsfield credits.
You can now choose when the plan card starts a long task by itself. Go to Settings › Agents › Show the plan before long tasks, and set "Start the work": after 10 s, 30 s, 1 min or 5 min, after a time that you enter, or only when you press Start. When the plan waits for you, you get a notification. The time applies to all agents on this computer. The default is still 10 s.
When a plan card waits a minute or more, its Start button counts down in minutes, such as "Start · 4:59".
You can choose the provider account when you create an agent. When a provider, such as Claude, has more than one account connected, the model picker of the new-agent form shows each account. Before, a new agent always started on the provider's own sign-in, and you had to change the account after creation.
You can now change how many messages and routine runs local scripts can send each agent in one hour. Go to Settings › Agents › Local scripts, and set "Requests an hour": 30, 60, 120, 300 or 1,000. The limit applies to all agents on this computer. The default is still 30. Before, a voice agent or another script that sends many messages stopped at 30 requests an hour.
The Usage page has a Skill tab in its Breakdown. It shows how many times your agents loaded each skill, the estimated tokens and cost of the skill text, its share of the tokens of all skill loads, and how many agents used it. OpenBot counts skill loads from this version on.

Improved

The web client reads the latest messages of the 3 agents that you opened last, after the open agent loads. The first switch to one of them after a page load shows its messages at once.
When you mark a chat read or unread on a team server, the web clients, phones and joined desktops of other members no longer load the chat again. Only your own devices do.
The ChatGPT code sign-in now tells you to turn on device code sign-in for Codex in the ChatGPT security settings first. For a work account, the workspace admin turns it on. Before, the sign-in waited with no hint when this setting was off.
When you ask in a channel for a new channel or a change to another channel, the agent now makes it there. Before, you had to ask again in the agent's own chat. Requests from routines, agents and team members on other computers still cannot do this.
OpenBot now measures how long a hosted server takes to start and how long the app takes to connect to a remote server, so slow starts and slow connection steps can be found and fixed. Only times, fixed step names and counts are sent, never a server ID, address or key. The Privacy page lists the new values.
Show a short line of context under an agent question. The ask_user tool can now send it, and a plugin form that asks for a value shows its request message under the first field.
The chat color picker has 11 new preset colors. They have the same lightness, so each bubble tint has the same weight. A color that you saved before stays as it is.
In the dark theme, a bubble with a chat color shows more of the color. Before, the tint was dull.
The Mobile tab of the changelog shows "Skipped" for a store that will not ship a version, because a newer version replaced it. Before, that store showed "In review" with no end.
In the web app, server owners now see Company sign-in in Members. Setup needs the desktop app, so Set up shows a message with a link to download it.
When product analytics is on, OpenBot now records each webhook request that reaches your computer: the kind of owner (agent, channel or workflow), the result, the HTTP status, the body size and the answer time, at most 60 each minute. It never records the URL, the secret, the event type or the body. Routine run events now also tell a run that a webhook started.
Cards where an agent asks you for an answer, such as a form for an API key, a choice or a confirmation, now look like the agent's questions. The buttons, including Skip, are at the bottom right, and the "Needs your answer" label is gone.
In the API key form, the choice of who can use the key is now a small menu next to the submit button.
The sidebar badge of an agent that waits for you is now felt: an amber felt ball with a cream yarn icon, with the same ring as the other felt badges.
Remove the "development preview" and "early stage" text from the website, the welcome email, the README and the security policy. OpenBot is ready for production use and gets a new release almost every day.

Fixed

An agent chat no longer shows an old reply at the bottom, below newer routine runs. Before, when routine runs pushed the latest reply out of the loaded page, a routine run moved that reply to the end of the open chat until you opened the chat again.
Find npx and other MCP server commands on your PATH when your shell profile is slow, exits with an error, or has an alias of the same name. Before, the Apple apps server and other MCP servers showed "Command not found: npx" until you typed the full path.
On Linux, show how to fix a real-time CPU limit of 0 when the system stops a provider because of it: install and start rtkit, then restart xdg-desktop-portal and OpenBot. Before, the provider showed only "stopped before it answered (signal SIGKILL)", and each Retry removed the downloaded provider and downloaded it again. While the limit stays at 0, a provider update also does not download a new copy that the system would stop in the same way.
Claude stays signed in after OpenBot restarts or updates. Before, the Claude sign-in check at start could use up the sign-in, and you had to connect your Claude subscription again.
Claude agents use less of the Claude 5-hour usage window. OpenBot now names itself to Claude, so its requests are no longer counted as Agent SDK requests, which the 5-hour window counts about twice as heavily.
A saved login or card now fills again after an agent runs a script on a site, as soon as the agent reloads the page, or closes it and opens the site in a new tab. Before, the fill stayed blocked until OpenBot restarted, and the message did not say how to restore it.
The browser can now fill saved logins from Proton Pass. Before, list_logins always failed with "The shared password vault could not be read", because OpenBot listed the logins without naming a vault. OpenBot now lists the logins of each vault that the access key can read.
On Windows and Linux, OpenBot now refuses Workspace only for a provider that it can only confine on macOS, such as a custom ACP agent, with the message to choose Full access. It also refuses a move of a Workspace only agent to such a provider. Before, the change was saved, and then each message to the agent failed until the user changed the access back.
In the sidebar, chats that scroll under a folder name now fade out, so you no longer see a chat cut in half below the folder name.
When you close the last tab of the browser in a conversation, the Browser pane now closes too, also from the full view. The next pane of the side panel shows, or the list of panes when no pane is open. Before, the side panel stayed open and showed "No browser tabs".
A teammate's question or answer now comes back to the side thread that asked, also when the teammate sends it without the id of the request. Before, it went to the agent's own chat, and the side thread waited without an answer.
When an agent showed "The agent could not continue. Try again.", the reason was not saved anywhere, so it could not be found later. The provider log (logs/providers/providers.log) now keeps the reason, the provider, the model and the turn of each agent error, with secrets removed.
Stop the console warnings that each shared button wrote when it appeared, for example the Back button in an agent question.
In a bubble with a chat color, code blocks, inline code, tables, file chips and cards now take the color of the bubble. Before, they stayed grey on the color.
Agents no longer read a "medium" effort for a model that has no effort setting, such as Cursor Composer. The model list that agents read now shows no effort for these models.
A chat with many answered agent questions opens faster. Before, each question card laid out the whole window again when the chat opened.
The Usage Breakdown no longer leaves empty space under a short table.
A thread that fails on a spent credit balance or quota now shows the provider's reason, for example "Credit balance is too low". Before, it showed "The provider gave no reason".
An agent that moved to a backup model because a credit balance was spent no longer goes back to that model after a restart or after 10 minutes. Before, its next turn failed again on the same balance.
A thread turn that fails on a provider error before it does work now moves an agent set to switch automatically to its backup model, as a failed chat turn does.
A file name in an agent answer, such as screenshot.png, now opens the file with that name that the agent attached in the same turn when the workspace has no file at that path. Before, the chip showed "Nothing exists at …", although the attached file was in the chat.
A Gemini agent on a second Google account can use the models of that account's plan. Before, the model menu showed the models of the first account for every account, so a Pro account had only the models of a free first account.
A hosted server whose stop the cloud provider cancelled starts OpenBot again. Before, the server ran with OpenBot stopped, and apps waited on "Starting" until the next stop.
GPT voice now works with every agent on this computer, also when the agent uses a Codex model. The voice gives each request to the agent's own provider and speaks the answer. A switch of the agent's model no longer stops the voice, and a start failure shows an error.
Text that you select in a side thread now gets Quote, Explain and the other selection actions of that thread. Before, a quote went to the agent's main chat box, and a sent instruction went to the main chat as a reply to a message that the main chat does not have.
A hosted or self-hosted server in an LXC container, for example on Proxmox, installs updates again. The update no longer stops when the container cannot load the AppArmor profile of OpenBot.
When an agent tries to remove an API key that is in the Vault, it now tells you to delete the key in Server settings > Vault. Before, it said that the key was not found, and the key stayed.
A message that you steer into a running side thread now shows in that thread at once. Before, it showed only after a restart of the app.
Stop asking for automatic compactions of a ChatGPT (Codex), Muse, Pi or Antigravity chat after 2 compactions in a row fail or time out. Before, each failed compaction let the next turn ask for another one. One chat compacted every 19 seconds for 7 hours overnight and used up the plan limit. OpenBot now shows a notice and stops. "Compact now" and "New chat" stay available. A compaction that completes, or a chat that falls below 80% of its window, starts the count again.
A message sent after OpenBot stopped a compaction that Codex did not stop no longer asks for a new compaction at once. It waits for the running compaction to end.
Usage no longer shows a cost of US$ 0.00 for OpenCode Go models. From OpenCode 2.2.0, OpenCode reports what the Go subscription charged, which is $0 for each turn. These turns now show "Unavailable" and count as unpriced. Turns recorded before this change keep their $0 cost.
In a long agent chat, older messages that you scrolled up to load now stay loaded. Before, a provider status update, such as a sign-in check or a provider that stopped when idle, reloaded the latest messages, and the chat jumped back. Older pages then loaded again, so the chat moved in a loop (#5259).
"Mark all read" in the sidebar now also reads the side threads of an agent, such as the thread of a routine. Before, the unread count of the agent stayed after "Mark all read" and after the agent's chat opened.
A routine run that ran in the routine's own thread now opens in that thread from the routine history. Before, the chat showed "This message is no longer available."
An error that only the app shows in a chat no longer has a Delete action. Before, Delete failed with "This message is not in the chat."
A file chip that joins a file name to a folder in the same message now opens the file as the agent wrote it, from the agent workspace, when the folder does not have it. Before, test/p1.txt after a Shared folder failed with "Nothing exists at …".
A file chip with only a file name now finds a file that a Pi agent read, wrote or edited in the same thread, also outside its workspace, as it does for Claude and ACP agents.
On Linux with Wayland, the phone now gets agent notifications when you are away from the computer. Before, OpenBot could not see that the computer had no input, so it kept the phone quiet all the time.
GitHub Copilot with no signed-in account shows Sign in. Before, it showed "ACP CLI did not advertise any ACP models".
Show full option names in an agent's choice card. Before, a long option description took most of the row, and the option name broke after a few letters.
A file chip with only a file name, such as main.log, now finds a file that a Claude or Codex agent read with a shell command (cat, tail, sed -n) in the same thread, also outside its workspace. Before, the chip failed with "Nothing exists at main.log in the agent workspace".
On macOS, a click on the server name in the title bar opens the server menu while Settings is open. Before, the click moved the window and the menu did not open.
A Grok agent on a newer model, such as Grok 4.7, now keeps that model when Grok starts with an expired sign-in. Before, Grok listed only its built-in models for a moment, and OpenBot moved the agent to Grok 4.6 for good.
Show a Cloud server as stopping, not running, when its provider cannot save its disk at a stop, and log it for the operator. Before, the Worker read the server as running and the fault stayed hidden.

Nov 5, 2026

Yaji 0.62.0

New

Add Download to the Hermes Agent row in Providers, on a server and on a Mac or Linux computer. OpenBot runs the official Hermes installer, shows its progress, and then shows Connect for the Nous Portal code. Windows keeps the install guide.
Connect TickTick from the Apps marketplace with browser sign-in to see and add tasks, organize lists and tags, and track habits and focus records.
A thread tab and its row in the thread menu show a blue dot when the agent wrote in the thread and you did not read it. The agent's unread count in the sidebar now includes these messages. Open the thread to read it.
Type @ in an agent's chat to tag one of its browser tabs or side threads. The agent gets the tab id and can use the tab, also from a side thread. For a thread, it gets the state and the start of the last answer, and it can read more of that thread.
A provider page in Settings → Providers now has a Delete button at the end. It turns the provider off and deletes the CLI that OpenBot downloaded for it. Accounts, sign-ins and agents stay. Before, you had to turn the provider off first and then remove the download.
Settings > Voice offers two more local dictation models. Parakeet Unified EN 0.6B is English only, as fast as Parakeet TDT, and writes punctuation. Cohere Transcribe knows 14 languages, and you must set the language. It is about three times slower, it is a 2.9 GB download, and it uses about 4 GB of memory.
Let Mark unread on a message work in the chats of a server that you joined. Before, only an agent on this computer had it. The server must have the latest version.
Owners and admins of this computer's server can add and change Telegram and Slack workflow triggers and outputs from the mobile app.
An agent can send files, such as an image that you attached, to a WhatsApp, Slack, Discord or Telegram chat when it posts there on its own. The files must be in the agent's workspace or the OpenBot shared directory. On WhatsApp, JPEG, PNG and WebP images arrive as photos that the recipient can save.
Choose the reasoning efforts of each custom provider model, such as Low, High and Max. OpenCode and Pi show only those efforts and send the one you select. Pi custom models can now reason. With no effort selected, OpenCode keeps Low, Medium and High, and Pi does not reason.
Let agents create a repository with the GitHub connection. The OpenBotGit app now asks for the Administration permission, because GitHub needs it to create a repository. An owner of each account approves the request in GitHub. Until then, the agent explains why creation failed and what to do. The openbotgit[bot] tokens never get this permission, so a user with push access cannot change repository settings through them.
Connect Proton Pass in the Marketplace, on this computer or on a server. Click Connect and sign in to Proton in your browser. OpenBot then makes a "Shared with OpenBot" vault and a read-only access key for one year, and keeps only that key. The browser can fill the logins of that vault. You can also paste an access key that you made. The Proton Pass CLI needs a paid Proton plan.
Give a named browser profile its own proxy. In Server settings > Browser profiles, add a proxy with one address, such as socks5://user:password@host:1080 or http://host:3128, and choose it in the Proxy menu of a profile. A user name and a password are optional, for SOCKS5 too. A profile without a proxy uses the proxy settings of the computer that runs OpenBot. A badge shows the proxy of each profile, and it shows when the proxy does not answer. OpenBot keeps the password encrypted and shows it only as a mask. When the proxy cannot be set, the tab does not open, so pages never go out without the proxy. Owners and admins of a joined server can do the same from the desktop app and the web. The address can also be user:password@host:port or host:port:user:password. The agent's profile select and the toolbar label of a tab show the proxy. A page that does not open names the proxy, and tells when the proxy refused the sign-in. Before a proxy change closes open tabs, OpenBot asks. A save that does not change the address keeps the tabs open. When two proxies have the same host, their names also show the port.
Use your own computer as the proxy of a browser profile on a joined server. In the Proxy menu of a profile, select "Use your computer…", or select Share in the Proxies section. The pages of that profile then go out through the internet connection of your computer, while OpenBot is open on it. Your computer opens only public internet addresses, so the server cannot reach your local network. Select Stop sharing to stop. Owners and admins can share.
After two days and ten messages, a small toast asks you to post about OpenBot on X, Threads, LinkedIn, Reddit, Instagram, TikTok or Facebook, and to follow OpenBot there. If you close it, it comes back after 7 days and then after 30 days, and then never. A post or a follow stops it.
In a channel, an agent can reply to a message of another member to ask it a short question or to agree on something. The reply shows in the chat as a reply to that message. The other member gets one turn to answer it, and the agent continues with that answer before it answers you. A request allows two of these replies, so agents do not talk without end.

Improved

The Browser pane of the side panel no longer has a second row with only a + button. New browser tab is now the first item of the + menu in the side panel header.
When a Marketplace search or filter finds nothing, or a tab is empty, a felt plush sits above the text and fishes. With reduced motion, it is a still image.
The @ list shows agents, tabs and threads in sections. The highlight follows the arrow keys, and the matched text is marked.
The daily resource summary now tells when the longest main-process delay came (start, wake from sleep, first hour or later) and counts helper process exits by process type. The local trace also records each main-process delay of 1 second or more. The summary holds no names or paths.
Usage analytics now include how long OpenBot takes to start and which startup step is slowest, how long a turn waits in the queue and until the first output, and, in the daily summary, the number of slow tasks and the input delay of the app windows. These are rounded times and counts only. The app keeps the time of its last start in a local file to tell a first start after boot.
Usage reports of a failed turn now include a fixed code for the reason, such as an expired sign-in, a usage limit or a context that is too long. The reports never include the error text.
Product analytics now counts how often OpenBot ends without a clean quit: a crash, a force quit, a kill or a power loss. At the next start, OpenBot sends one event with the phase, the previous app version, a coarse uptime class, the number of new local crash dumps, and whether the computer slept or started again. Crash dumps stay on the computer. The analytics setting applies.
Local dictation with Parakeet or Bark and replies read aloud use less memory. After a long turn, the dictation process holds about 230 to 500 MB less, and the read-aloud process about 110 MB less. Whisper keeps its settings, because they make it faster.
Local dictation uses half of the logical processors, from 4 to 6 threads: 6 on a computer with 12 or more, and 4 on a computer with 9 or fewer. With the memory change, long turns are as fast as before or a little faster.
Underdog Bark 3.0 Plus is no longer offered, because Parakeet Unified EN is as fast and makes fewer errors. If you use Bark or have downloaded it, it stays in Settings > Voice and continues to work.
When an agent asks for an API key in the chat form, the form now saves the key in the Vault for all agents by default. Select "Only this agent" to keep the key for the asking agent only. Other agents then use the Vault key and do not ask again. The asking agent's own copy of the key is deleted, because it would hide the Vault key. Another agent that saved its own copy before keeps that copy, and it does not get Vault changes until that copy is removed. Without a Vault, the form saves the key for the asking agent, as before.
The Model picker in Brain of the agent settings now shows Effort and the Fast switch in its menu, the same as the model picker in the chat header. Both pickers use the same settings. The separate Reasoning row is removed.
A side thread shows its queued messages in the same queue as the agent's own chat. You can steer, delete, reorder and edit a queued message there. Before, the thread showed only a plain line of text for each message.
Show agent questions in one chat bubble, one question at a time. The next question moves in after you pick an answer, Back returns to an earlier question, and dots show the progress. Each option has a letter key, and "Something else" takes a typed answer. When you send the answers, the bubble shows a summary of each answer.
Run agent routines in the agent's chat again, as before 0.61. A new switch in the routine settings, "Run in a separate thread", keeps the runs of one routine in their own thread. Routines that you made before this release run in the chat.
Icons have a slightly heavier line, so they are easier to read. Larger icons, such as the icons of the page rail, dialogs, notices, app tiles and workflow tools, have a lighter line, so all icons look the same weight.
All icons have the same line. Some icons, such as the lock on an agent row and the More and Help buttons on the page rail, had a heavier line than the icons beside them.
Icons no longer fill in or change their line when you point to them or select them. They only turn white. This applies to menus, the page rail, the pin of a pinned page and the pressed tool on a workflow canvas.

Fixed

OpenBot shows fewer "Not connecting" errors when the account service database is busy for a short time. The account service now tries a failed read one more time.
Codex agents with many MCP servers use less context. Before, each request sent the full tool list of each MCP server. Now Codex finds these tools when the agent needs them. Computer Use tools stay in each request.
A sent image stays in place when the server stores the message. Before, the image went blank, became square for a moment, and the chat jumped.
Hermes Agent installs on a self-hosted server and in the Docker image. Before, its installer stopped because the server had no libatomic1, and the service user cannot install packages. The server guide now tells how to install Hermes for the OpenBot service user.
Connect shows for Hermes Agent, Kimi and GitHub Copilot on a server, in the app and on the web. The server signed them in with a code, but only the iPhone app offered it.
OpenBot now tells every provider that tools are unavailable when it writes a channel summary, a title or an agent profile. Before, only Codex got this instruction. With OpenCode and other providers, a model could try a tool, for example to count the characters of a summary, and the generation failed. The channel summary then stayed at an old version.
A hosted server downloads its own Claude or ChatGPT CLI when the CLI of the server image is too old. Before, the provider showed as too old, and it looked uninstalled.
ChatGPT no longer fails for 15 minutes after a slow first start. OpenBot gives the first start more time, and an update no longer reports "could not update the ChatGPT CLI" when ChatGPT only prepares its local history.
OpenBot no longer stays on "Connecting to OpenBot" at start. Before, when the account check finished while the window was still loading, the window did not get the result and showed the connecting screen until a restart.
The gap between the agent's chat tab and its thread tabs in the chat header is now the same as the gap between two thread tabs.
A joined server loads its agents and chats at app start without a click on Retry. Before, when the server connected while the window was still loading, the window kept the server as offline and showed the loading screen until Retry.
When a scheduled, script or webhook routine run has nothing to report, the agents linked after it on the Routines canvas are now skipped. Before, each of them started with "(no answer text)" as its input.
Show the models of a custom ACP agent that reports model values with quotes, such as DeepSeek Harness. Before, OpenBot dropped all such models, so the agent had no models and you could not select it.
The app pauses less when it opens a chat with routine cards or many messages. Before, it laid out the page again for each routine card, and again while the message rows loaded.
Codex agents no longer say that the browser or another OpenBot tool is unavailable when an exec cell does not list it in ALL_TOOLS. OpenBot's tools are direct tools that exec does not list, and the agent instructions now say so.
Scrolling a long chat no longer stops for a moment. Before, three menu style rules made the app restyle the full chat when the scroll shadow at the top or bottom changed, or when the "scroll to latest" button showed. In a chat with many code blocks, each stop took more than half a second. Chats with many code blocks or formulas also open faster.
The first shared-table call of an agent after OpenBot starts no longer fails as a slow statement. Before, the start of the database process counted against the 5 second limit of that statement, and a slow Windows computer often went past it. The start now has its own limit, and OpenBot starts the process early on a computer that already has shared tables.
OpenBot runs a turn again on the backup model when OpenCode, Pi or Antigravity stops it at a rate limit, a usage limit or an expired sign-in. Before, OpenBot did not identify some of these errors and showed the error without a try on the backup model.
Devin and other agents that write Rust log records no longer show an info or warning record as a provider error. Before, a record such as INFO … request failed, will retry from Devin CLI showed as an error and wrote "A provider failed." to the log. Error records still show.
The WhatsApp Orchestrator answers the owner's own requests in the "Message yourself" chat of a linked number. Before, it read those messages as from another person and refused requests for the owner's own data, such as their latest emails. Other people keep the same limits.
Pi now lists Cloudflare Workers AI models, such as @cf/moonshotai/kimi-k2.6. Before, OpenBot dropped each Pi model whose id has a character such as @, so a Pi that is signed in only to Cloudflare Workers AI showed no models.
You can now change the name and logo of a server that you created on this computer. Before, the app showed "Sign in to OpenBot first." also when you were signed in.
OpenBot sends sign-in codes, invitations and other emails through a second provider when the main mail server reaches its hourly limit. Before, sign-in stopped with "OpenBot cannot send more email right now" until the limit reset.
The agent that answers Slack, Discord, Telegram and WhatsApp now saves a lasting fact that a person gives, such as the repository or account to use. Before, it kept the fact only in that conversation, so a new conversation or a restart lost it.
An agent answer that names an image file, such as image.png, no longer shows an older picture with the same name from an earlier message. The inline preview now shows only the file at the named path. The file chip still opens the file.
The "Service unavailable" screen at start now goes away by itself when the account service answers again, for example when the network is ready only after the first 30 seconds. Before, you had to click Try again. The code on the screen now also names the failure, such as no network or a service error, not only auth/unavailable.
macOS keeps one Dynamic Island on each display. Before, after the displays slept or changed, the island of one display could stay on a different display, under the island of that display.
When a teammate could not start or finish a request from another agent, the agent that asked was not told and could stay idle. Now it gets a note with the reason, so it can tell you what failed.
A plan that closes no longer moves a chat that you scrolled up. Before, in some cases the chat jumped to the latest message and then slid back down while the plan closed.
A side thread stays at the latest message while the agent writes a fast reply. Before, the thread could stop following the new text with no scroll by the user.
The app pauses less when it opens a chat with code previews, and when older messages load while you read. Before, each code preview laid out the page when it showed, and each loaded page of messages laid out the page again.
OpenBot fills a Bitwarden login on every page of its site when the login uses the default or "Base domain" match rule, as Bitwarden does. A login saved for myaccount.google.com now also fills on accounts.google.com, and a saved address with no https:// now fills. Before, OpenBot found the login only on the same address, so the agent asked for the password and the code again at each step.
The OpenRouter and DeepSeek model pickers show models that the provider added after the bundled Pi release, such as OpenRouter's stealth model Crouton Alpha. Before, the picker showed only the models that Pi had at its release, or at its last full refresh.
The Vault imports the user name of a Proton Pass login that has a username and no email. Before, the CSV import left the user name empty. The import text now names Proton Pass.
The window no longer stops to respond when the browser panel changes to the active tab of the agent.
OpenBot no longer asks for automatic compactions of a ChatGPT (Codex), Muse, Pi or Antigravity chat that stays at 80% or more of its context window after each compaction. Before, OpenBot stopped only at 90% or more, so a chat that stayed at about 84% compacted again and again and used up the plan limit. Now OpenBot stops after 2 such compactions and shows the notice that the compaction does not make the chat smaller. "Compact now" and "New chat" stay available.
The loading screen animation stops after the app opens. Before, the plushes of the closed loading screen kept their animations, so the window used CPU for each display frame also when nothing moved.
A routine run and a chat in a routine thread have Computer Use again when it is on for the agent. Since 0.61.0, each routine runs in its own thread, and these threads did not get the Computer Use tools, so a scheduled task that uses a desktop app, such as Messages, could not run.
When the orchestrator of a Slack, Discord, Telegram or WhatsApp conversation gave work to a teammate, and the teammate asked a question, the question waited in OpenBot and the conversation looked stuck. Now the teammate decides on its own. When only the OpenBot user can help, for example with a sign-in or a key, its answer in the conversation says what to do in OpenBot.
On macOS, the transparent window is clear in all of the window. Before, only the frame and the card were clear. The message bubbles (also bubbles with a chat colour), tables, code blocks, the message box, the Settings cards, the Tasks message box and the Workflows canvas stayed solid. Menus, dialogs and tooltips stay solid, so their text is easy to read.
Agents in the Docker image find the programs in ~/.local/bin, for example a Python that uv installs. These programs are in the data volume, so they stay after an update. Before, only /data stayed, and on an instance with an empty data folder agents did not find the programs in it.
OpenBot uses less energy when it is open with no task. Before, the app woke the processor many times each second while it did nothing. The main process measured its own response time all the time, the Computer Use highlight checked the screen 30 times each second for the whole session, and the agent avatar and the sample turn of an empty chat kept moving when another app had the focus. These now rest until a window has the focus or an agent starts a turn.
In menus, icons are grey at rest, as on the page rail. The icon of the row under the pointer or the keyboard turns white with its label.
In the settings navigation and on the composer, icons are grey at rest and turn white when you point to them or move to them with the keyboard. The icon of the open section, and a composer button while its menu is open, stay white. The warning icon of a marketplace app keeps its color.
The icon of a disabled menu item stays grey when you point to it.
Icons in setting rows, such as the 1Password steps and a background process with no agent, have the size of the other large icons. Their titles now align with the titles of agent rows.
The file icon in the agent instructions bar and the icon of a custom provider now have the correct size.

Nov 4, 2026

Yaji 0.61.0

New

A GitHub trigger of a workflow can watch more than one repository. An event from any of them starts a run, and a GitHub output set to "Same as trigger" acts on the repository of that event. Workflows that watch one repository do not change.
Devin is a new provider. OpenBot downloads the Devin CLI from More providers, and you sign in with your Devin account in the browser. Devin agents use the models of your Devin plan.
The agent colors have lime and indigo again. Agents that had lime show lime, not green, and agents that had indigo show indigo, not violet.
Agents can use more colors: sky blue, navy, pink, wine, brown, sand and gray. An agent keeps its exact color, and a joined server or the web client shows it when the server has the new optional agent-colors-v1 capability. Older apps and servers show the nearest released color.
In a channel, $ now lists the skills of the agents in that channel, as it does in a chat.
Agents can create and change channels from their own chats and from threads that you open. An agent can set the name, title, instructions, members and lead, archive or restore a channel, and give it a picture: an icon on a colour or an image file.
You can give an agent the flat avatar style again. Open the avatar editor in Agent settings and select Flat next to Shape. The shape, eyes and color stay the same, and Plush changes it back. The iPhone app shows a flat avatar as a plush avatar.
Webhook routines accept Standard Webhooks requests, such as Composio, Svix, Clerk, and Resend send. To use the signing secret that the sender made, select Use the sender's secret in the routine editor or on the chat card. Filters can read the complete request, for example /metadata/trigger_slug for a Composio trigger.
WhatsApp shows in Marketplace › Apps on a joined server. An owner or admin links the server's WhatsApp number from their own computer or the browser client: the server makes the QR code or pairing code, and your screen shows it. The server must also run this version.
You can delete a message in a chat. Right-click the message and select Delete. This works in agent chats, side tasks and channels, on this computer and on servers that you joined. The message goes from the chat, its history and search for everyone. The agent can still remember it. On a server, members delete their own messages; admins, and members who can manage the agent, delete any message.
One server can have a custom limit of active hosted sites, such as for an enterprise customer. The limit replaces the site limit of the server's plan, from 1 to 1000 sites. OpenBot staff set it.
A commercial license gives 3 hosted sites for each user. The servers of the account with no plan share these sites.
A workflow Condition step has a model and effort picker, as an LLM step has. A new Condition step starts with the model of a new LLM step. A Condition step saved before shows "Auto": the server chooses its model until you choose one.
Chat colors. Pick a chat color for an agent in its settings, and one for yourself in Settings → General. The color tints the chat bubbles, and everyone on the server sees it. The avatar keeps its own color.
The website has the OpenBot Contributor License Agreement at openbot.run/cla. You can read it on the page or download the PDF to fill in, sign and email to licenses@openbot.run. The site footer links to it.
A guide that compares chat history with saved workspace files in OpenBot, using a recurring research task and plain explanations of model changes and agent handoffs.
The Usage page shows the CPU and memory use of the server to owners and admins. A click shows the details and the disk use. The disk row opens Storage.
When you scroll away from a subagent card while its subagents work, its header now stays under the chat header with the progress, the current step and the time. Click it to go back to the card.
GitHub Copilot is a new provider. OpenBot downloads the GitHub Copilot CLI from More providers, and you sign in with GitHub in the browser or with a device code. Copilot agents use the models of your Copilot plan, and Disconnect signs the CLI out.
Agents can make, change, delete and run workflows. An agent can add, change, move, connect and remove steps on the canvas, start a run and read its result. The tools load only when an agent searches for them, so other turns do not get longer. Only your own message to an agent, in its chat or in a thread you opened, can make it change, delete or run a workflow; a routine run or a team member's message cannot. Before an agent deletes a workflow, OpenBot asks you to approve, also when the agent has Full access.
When an agent builds or changes a workflow that is open on the Workflows page, you see its cursor go to each step, and the steps and connections show one after the other. A workflow that an agent makes while the page is open opens on its own.
A workflow that an agent builds has a clean layout: the steps do not touch, each connection shows, and a condition's "No" step is below its "Yes" step. Ask an agent to tidy a workflow to lay it out again.
You can talk to a Codex agent with Codex live voice. Turn on Codex live voice in Settings → Voice, then open the Voice pane of a Codex agent on this computer. The agent hears you and answers aloud in real time. What you say, what the voice says, and the work of the agent show in its chat. The conversation continues when you open another tab or another chat, and stops when you close the Voice tab. Live voice uses your Codex sign-in, and your voice goes to OpenAI. It is off by default.
You can choose the voice of Codex live voice in Settings → Voice. "Automatic" lets OpenAI choose.
In a channel, you can ask the lead to drop earlier work, for example when a newer request replaces it. The lead cancels that whole run with its running and queued sub-tasks, and the chat shows a note that names who asked.
Each row of "Waiting for sub-tasks" in a channel now has Stop. A stopped sub-task then offers Continue, Reassign and Cancel.
Connect a Coinbase or Robinhood account from the Apps marketplace. Agents can read balances, positions and prices, and place orders.
Agents can save an API key or a website password that you paste in the chat to the Vault. Before, the agent asked you to type the key again in a form. A pasted key is now available to every agent, and the chat shows a note when an agent adds or changes a Vault item. This does not work in Slack, Discord, Telegram or WhatsApp chats.
Connect RevenueCat from the Apps marketplace with a secret API key to manage products, offerings and paywalls, look up customers, and read revenue charts.

Improved

Each routine of an agent now runs in its own thread. A routine no longer waits for other routines or for the agent's chat, and the chat does not wait for a routine. The answer of a run shows in the routine's thread, and the chat shows the run with an "Open thread" link. A Test run opens the routine's thread. Before, all routine runs waited in one queue in the agent's chat.
A routine run in its own thread still shows a notification, sends a phone notification and plays the completion sound. The notification shows the answer and opens the routine's thread.
A routine run no longer uses Computer Use: only the agent's chat controls the screen, so a routine and the chat never control it at the same time. A run that needs Computer Use says so.
An agent with no thread limit of its own now runs up to 4 threads at the same time. Before, the default was 2.
The provider list in onboarding has at most three plan rows, also when you have many providers. The others are in "More providers". A new computer now shows ChatGPT, Claude and Gemini.
In settings, the section title, or the path to an open page such as Providers › ChatGPT, is now in the top row of the window, as on the other pages.
On an open settings page, such as Providers › ChatGPT, Back in the top row goes to the list of the section before it goes to an earlier chat.
Agents ask for a password for a script or a local service, such as an IMAP or Proton Mail Bridge password, with the secure key request. They do not ask for it in chat.
When an agent works, its sidebar badge shows three felt balls that hop in a wave, in place of the grey dots. With reduced motion on, the balls stay still.
The app demo on the home page opens on the workspace at once. Before, the loading screen with the falling plushes showed first.
The servers of an account with no plan and no license share 1 hosted site. Before, each server had 1. No site is deleted: a server above the limit keeps its sites until they expire.
An account can have at most 300 active hosted sites on its servers.
The privacy policy has a new Google User Data section. It tells which Google data OpenBot reads, how it uses, stores, shares and deletes this data, and how OpenBot obeys the Google API Services User Data Policy.
Agents show a confirm card with real actions, such as Send, Edit and Keep draft, after a draft or a plan for a send, a post, a delete or a payment. Before, they often asked in a sentence, and you typed the answer. openbot.ask_ui is now always available to agents, without a tool search (#3922).
New contributors sign the OpenBot Contributor License Agreement instead of the Contributor Assignment Agreement. You keep the copyright in your contributions. You sign the PDF and send it by email; a ticked box or a comment on the pull request no longer signs. A signature of the assignment agreement still counts.
After the network comes back, the web client reads the agent list, channels, account usage and side threads one time, not two times. A reconnect sends about 10 fewer requests.
Settings → Voice has four groups: Dictation (models and language), Spoken replies, Live voice and Voice server. When Codex does not use a ChatGPT sign-in, Live voice shows Connect, which opens Providers. When OpenAI refuses the sign-in, the Voice pane says to sign in to Codex, not a raw error.
A voice conversation continues when you open another tab of the side panel, such as agent settings or the browser.
When a search finds nothing, the lilac plush shows above "No results". Its eyes follow the caret while you type. When you stop, it looks for a result and tilts its head. With reduced motion on, the plush stays still.
Agents finish more of a task before they stop. They wait in the same turn for a build or deploy of up to about 15 minutes, try other ways when a step fails, do work that their own tools cover instead of passing it to a teammate first, and finish the rest of a task when a teammate returns only part of it. Before, an agent often stopped with an unchecked deploy or with a teammate's partial result, and the user had to tell it to continue.
In a channel, an agent's steps row no longer stands above its name. When you hover an answer, the copy and reply actions and the steps show beside the bubble, or under a long answer. The steps open the full trace of the turn, named for the agent that answered.
The sidebar badges are now felt. An unread count of an agent, a channel or a person shows cream felt digits on a plum felt pill, a failed routine shows a felt X, and the "+N" count of a channel uses the same felt pill. Each badge has a thicker ring in the colour of its row.
An unread count and a failed X now fade in quickly when they show. An unread count bounces and rolls its digits when it changes, and shrinks away when you read the chat. With reduced motion, the badge changes with no movement.
A pinned channel now shows its unread count, as a pinned agent does.

Fixed

The line under the search field of a workflow menu no longer shows in the focus colour each time the menu opens.
A long label or repository name ends with "…" in its tag on a workflow card. Before, it went past the edge of the card.
In the new agent setup, you can click every color. Before, the large avatar covered the colors in the middle of the row, so a click there did nothing.
A channel assignment no longer shows as started while its agent's own chat starts a turn. Before, the assignment could be claimed in that moment and then stay "starting" with no turn until the chat turn ended. Now it stays queued until the agent is free.
The Vault and Roles pages of the server settings show in Polish when the app language is Polish. Before, these pages were in English.
In the agent avatar colors and the channel picture colors, a hovered color no longer covers the ring of the selected color. Each color tile is now as wide as its column.
On a Mac with a server on this computer, an empty second OpenBot window no longer opens next to the app.
Grok no longer shows a "Grok error" when Git is not installed, as on a Windows computer without Git. Grok continues without Git, so the message now goes only to the log.
An agent set to "Ask before switching" now continues its work on the first available backup model when a usage limit, such as the Claude 5-hour limit, stops it after it started to work. Before, the turn failed, and the agent did not ask or switch.
Settings › Agents shows in German, Spanish, French, Italian, Japanese, Polish, Brazilian Portuguese, Russian and Turkish. Before, its title, descriptions and summaries were in English.
When a Kimi turn fails, the error now shows the reason that Kimi gives, such as a rate limit. Before, it showed only "Internal error".
The felt plushes on the loading screen and in agent avatars no longer have a dark outline in the light theme. The edge of the fur now has the colour of the plush.
On a Linux computer with a Wayland session, OpenBot now starts when no monitor is on, for example after an automatic update. Before, it stopped with no window until a monitor came on. With no monitor, OpenBot now starts on XWayland, and the next start uses Wayland again.
In a shared chat, a member who joined a server after it started now shows with their profile name. Before, their messages showed an account id until the server restarted.
The Dynamic Island uses much less processor time when nothing changes. Before, the greeting motion and the logo blink made it draw on every screen refresh, all day. Now the greeting moves only while the pointer is on the island, and the logo draws only while it blinks.
On a remote server, the thinking of an agent no longer shows as cut-off messages above your message while a turn streams. It now stays in the thinking group of the turn.
The Usage panel now records the tokens and the cost of Grok agents. Before, it showed 0 tokens for Grok, and each Grok turn counted as a turn without usage. The cost is the cost that Grok CLI reports for each prompt.
On Windows, a channel lead can route messages again. Before, each message that the lead routed failed with "Routing failed. Choose a member or try again." when Windows still held the temporary folder of the routing turn. Side questions and new agent profiles use the same step.
A channel lead on Claude routes messages again when Claude signs in with a saved Anthropic API key and no claude.ai account. Before, the channel said that the lead could not route the message. Channel summaries, workflow steps, memory review and generated profiles on such a Claude setup work again too.
"Keep memories in this channel" in Channel settings now stays on. Before, the app did not save the setting, so agents saved what they learned in the channel as their own memories.
When an agent in a channel hands work to an agent that is not a member of the channel, the agent now learns which agent is missing, which members it can use, and that a person must add the missing agent. Before, it got a general message, tried the same handoff again, and the task stopped with no clear reason.
An MCP server whose command is a path to a file that does not exist now shows the "MCP server not started" error. Before, OpenBot gave the server to the provider, the provider could not start it, and the agent lost its tools with no message in the app.
In a channel, an OpenCode member no longer shows each note that it writes before a tool call as a message for a short time. The notes go into the steps above the member's answer, as before. The member's own chat still shows its text as it arrives.
On Windows, an agent that uses the Claude provider with many MCP servers can answer messages again. Before, each message failed with "The agent could not continue. Try again.", because the MCP servers made the command that starts Claude too long for Windows.
Agent faces in a group avatar have their outline again. In 0.60.0, the faces overlapped with no edge between them.
On Windows, you can now create a server on this computer. Before, the new server stopped at once, and the app showed "The server on this computer did not start."
A channel with a long history opens quickly again. Before, the app read every task of the channel again for each task, which kept the computer busy for seconds or minutes after start and after each channel message, so the channel took more than 10 seconds to open.
A channel no longer shows a member as working after its task stops or fails. Before, in a busy channel, "Working on it…" could stay until you opened the channel again.
An agent on Grok now moves to its backup model, or waits for the window, when Grok says that its free usage is spent (subscription:free-usage-exhausted). Before, a channel task stopped with the error, and each later usage limit moved the agent to the spent Grok model again.
Pi and Muse chats show your home folder in paths again when Computer Use is on. Before, the folder changed to •••, and a Shared file link in the chat could not open the file.
On Linux, main.log now records why OpenBot renders on the CPU, and why the GPU process stopped. Before, these lines went only to the terminal, so a GPU crash that turned off the GPU left no reason in the log.
A workflow Condition step now runs on a connected provider when ChatGPT is not connected and your default is ChatGPT with no model chosen. Before, the step failed with "Connect ChatGPT to continue". Another default provider, or a default model that you chose, stays the provider of the step.
A chat that is almost full no longer compacts again and again when the provider retries failed requests. Before, each turn compacted the chat, the agent did no work, and only a restart stopped it. When 2 compactions in a row do not make a chat smaller, OpenBot now shows "Compaction does not make this chat smaller" and stops compacting it automatically.
In a channel, when you quote a message and send only a mention, the agent now gets the quoted message with its author as part of the request. Before, the agent got only the mention and could answer about a different message.
A member of your server can open a channel again when the channel has an agent that the member has no access to. Before, the channel did not open for the member after that agent wrote in it. The member does not see the messages of that agent.
A member of your server does not see a channel when the member has no access to any agent in it. Before, the channel stayed in the list of the member.
Server Settings › Providers now shows the Default provider of a remote server to its owners and admins, on the desktop app and in the web client. Before, only this computer showed it, so you could set the provider of new agents on a remote server only from the iPhone app.
When you open the web client again, the open chat shows its loader until your server answers. Before, the chat showed the notice "Server actions will resume after reconnection" on each first connect.
Company sign-in, the agent colors, voice playback speed and the personal pricing text show in Arabic. Before, they showed in English.
A workflow with a title longer than 80 characters now hands its work to an agent. Before, the agent's conversation did not load ("Invalid conversation message.") and its queue stopped. Messages that a longer title made unreadable load again after the upgrade.
The Custom tab of the model picker no longer tells you to add an endpoint when your saved endpoints have no models in the list yet. Before, a save while OpenCode or Pi was running a task showed the endpoint count with the "Add an OpenAI-compatible endpoint" text. Now the tab says that the provider reads the models after its current task stops.
On the Usage page, the total cost uses a smaller font, so a large amount stays on one line. Before, the last digit could move to a second line.
The cost labels at the left of the Usage chart now show in full in all languages. Before, a long label such as "US$ 2.400,00" was cut.
The Totals row on the Usage page now hides the items that do not fit. Before, it could cut the last item in the middle.
A closed agent card is quieter. It says once, in one line, that it was answered, expired or closed, and a dangerous card loses its red border when it closes.
When an agent wrote a text, such as a draft reply, and then asked you a question about it in the same turn, the chat showed only the question. The text was in the closed steps of the turn. Now the text shows above the question until you answer it.
A side thread now shows the Sign in card when its agent needs you to sign in to an MCP server. Before, the thread showed only a line of text, and you could not sign in from it.
A Gemini message no longer shows "thinking" without end when Gemini sends nothing back. Before, you had to stop the message and send it again, sometimes more than once. Now OpenBot sends it again by itself after 3 minutes with no reply. If Gemini is silent again, OpenBot ends the message with an error, and you can send it again.
You can now sign in to an MCP server that follows the OpenID client registration rules. Before, some servers refused OpenBot with "web redirect_uris must only use the https scheme".
On a Windows computer with more than one monitor, the remote desktop now shows the monitor that you select. Before, it always showed the primary monitor, and the monitor names could be wrong.
You can invite people to a server that you made with New server on this computer. Before, the invitation stopped with "Sign in is required." Invited people join the server from their own devices, and their names show in the member list. They can connect while OpenBot runs on your computer and you are signed in.
When an OpenCode, Kimi, Cline or Hermes agent starts a subagent, the chat now shows the subagent card with its task, its state and its time. Before, these agents showed no card, so the agent looked idle while its subagents worked.
A Codex subagent card now shows the subagent's steps while it works: its commands, thoughts and messages. Before, the card showed only the task until the subagent ended.
An OpenCode subagent card now shows the subagent's reads, searches, commands, thoughts and messages, and the error when the subagent fails.
An agent whose subagents still run now shows as working in the sidebar, also when a background subagent runs on after the agent's turn. Before, the agent showed as idle.
When a subagent fails, OpenBot now tells you: a desktop notification when you are not in the app, and an alert when another chat is open. It uses the agent's notification settings.
Claude agents use fewer steps and tokens with the tools they find with a search. They no longer call a found tool directly and get a refusal, and a request for a workflow no longer loads Claude Code's own workflow skill.
A channel with a long history no longer stops when its summary fails. Before, each member's turn failed at start with "The provider could not generate a profile". Now the turn starts with the saved summary and the newest messages, and the next turn tries the summary again.
A queued message no longer waits until a restart after a turn end fails to save. Before, a context compaction that the turn end started to prepare held the agent, and each next message waited with the reason "compaction".
A Marketplace agent update no longer fails when one of its skills has local changes. Before, the update stopped with "This skill has local changes" and the agent stayed on the old version. Now the update asks about each changed skill that gets a new version: keep your changes, or replace them with the new version. A host that updates an agent for a joined server keeps the changed skills.
On Linux, OpenBot no longer asks you to sign in again after each restart or update when the system keyring does not keep its key, for example a locked keyring or one that keeps its items only in memory. Before, the saved sign-in and other secrets did not open after a restart. Now OpenBot finds this, keeps new secrets in its own key file that only your user can read, and the sign-in screen tells you why you must sign in one more time.
OpenBot is faster when it renders on the CPU, for example on a server, in a remote desktop, or on Linux after a GPU failure. Dialogs and panels then do not blur what is behind them, and agent avatars do not move.
After you stop or cancel a sub-task in a channel, the lead knows it. It no longer says that it still waits for that work.
When an older OpenBot starts over data that a newer version saved, it now installs the latest version and opens again. If it cannot update, it shows a download link and a troubleshooting guide. Before, it showed an error with only an OK button (#4969).
When OPENBOT_ANTIGRAVITY_PATH names a Gemini server file that is missing or not executable, OpenBot now says so and tells you to correct the path or remove the variable. Before, it said that Gemini was not downloaded and told you to download it, but Download refused while the variable was set.
When a sign-in server refuses to register OpenBot, such as RevenueCat's, the MCP server page now says so instead of showing "The client is not registered".

Nov 3, 2026

Yaji 0.60.0

New

Server settings has an Audit log. Owners and admins see who joined or left, who changed a member's role or a custom role, and who changed the access of each agent, with a filter and a search.
You can turn on Fast mode for a Codex agent on a joined server, from the model picker, as on this computer. The server must run this version or a newer one.
An owner or admin can change more settings of an agent on a joined server from the desktop app: Computer Use, "Agent starts new chats", the agent's folders and excluded folders, and Publish. To add a folder, you pick it from a list of the folders on the server.
An owner or admin can edit the instructions file of an agent on a joined server. The server must run this version or a newer one.
You can compact the context of a chat or side task with an agent on a joined server. The server must run this version or a newer one.
When an agent asks you to add a card, an address or a login to the Vault, the phone app shows the same card as the desktop app. The server must run this version or a newer one. Add opens the Vault form with the values that the agent filled in, and you type the secret values there.
A workflow that a Discord mention starts now keeps the files of the message and of the message that it replies to. Agent steps get the name, type, size and address of each file. A GitHub output that makes a new issue or a comment copies the files to the openbot-attachments branch of the repository and shows them in the text: images show in the issue, other files are links. A file larger than 25 MB, a file after the first 50 MB, or a file that cannot be copied, is linked on Discord, where its address expires after about a day.
A file that downloads from a joined server now shows its progress. The download button becomes a ring with the bytes received below the file name, and a click on the ring stops the download. A failed download shows "Download failed" and a retry button. Before, nothing changed on screen until the whole file arrived.
OpenBot can take its colors from a theme palette file in ~/.config/OpenBot/themes/, such as the one that an Omarchy theme writes. Choose the palette under Settings > General > Theme. The window follows each change to the file, with no restart.
An agent can give work to a channel from your direct chat. Ask it to send a task to a channel, and the channel members do the work there.
"Send to channel" on the message menu of an agent chat opens a channel with the message text in its composer.
Your phones and other computers can now open the servers that OpenBot runs on this computer beside its own server, and get their notifications. While OpenBot is signed in, it adds each of these servers to your account. When you delete one, OpenBot removes it from your account.
Files that you attach to a chat show in the message box at once, before they reach the server. For a remote OpenBot, a ring shows how much of each file the server has. Before, a large file uploaded with no sign until it was ready.
A file that fails to upload stays in the message box with the reason and a Retry button. The other files of the same pick still arrive. Send waits until you send the file again or remove it.
You can give a channel a picture. In channel settings, click the picture at the top to upload an image, or to choose one of 20 icons and a color. The picture shows in the sidebar, in pinned chats and in the channel header. The reset button shows the member faces again. You can change the picture only on the computer that hosts the channel; other devices still show the member faces. (#4274)
A channel setting, "Keep memories in this channel". When it is on, an agent saves what it learns in the channel as a channel memory, and its other chats do not get it (#4332).
Each provider page in Server settings → Providers has a Models section. Turn a model off to remove it from the model picker. A search and Hide all help with long lists, such as the models of OpenCode. Reset to default undoes your changes. Agents keep the models that they use.
The model picker shows the newest models of each provider first. For example, Claude shows Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1. The older versions are under Older models at the end of the list, and a search finds them too. A new release shows at once, with a New label for its first 7 days, also after you change the list.
Side threads have Reply on each message, as the agent's own chat has. The message box shows the quoted message, and the sent message shows the quote above it. A click on the quote goes to the quoted message. A joined server needs this version for Reply.
WhatsApp: link a WhatsApp number to this computer under Marketplace → Apps → WhatsApp, with a QR code or a pairing code, and talk to your agents from WhatsApp. A WhatsApp Orchestrator answers each new conversation and asks the right teammate. With your personal number, only your "Message yourself" chat and the people and groups you allow can give work; with a number used only by OpenBot, other people get a code that you approve in OpenBot. In a group, the agents answer a mention of the number or a reply. Files go both ways. Approve or deny a request by replying 1 or 2. Messages go directly between this computer and WhatsApp; no OpenBot service sees them. WhatsApp does not support this kind of client officially and can block numbers that it sees as automated, so a separate number is safer.
When the disk of a hosted Starter or Standard server is almost full, OpenBot offers the next plan. The dialog compares the disk, memory and price of the two plans, and opens the plan change in billing. It shows once when space is low and once more when the disk is full.
In Discord, the agent now reads embeds, such as the form of a ticket bot. Reply to the message with the embed and mention OpenBot. Link previews in earlier messages are context too. The agent gets the title, description, fields, footer and the links of the embed images. It does not get colors, icons or buttons.
Start a workflow when someone mentions OpenBot in a linked Telegram group, with filters for the chat and the text.
Send a workflow result to Telegram: reply to the message, post in a linked chat, or add a reaction.
Use OpenBot in Arabic. The layout mirrors from right to left when Arabic is the language.
Each chat paragraph shows in the direction of its text, in all languages. Arabic and Hebrew paragraphs read from right to left, and English words, links, numbers and code in them keep a readable order. Code blocks always read from left to right.
Start a workflow when someone mentions OpenBot in a Slack channel. Limit the trigger to one workspace, one channel, or messages that contain a text. A Slack output replies in the thread of the mention, posts in a channel, or adds a reaction to the mention. Only owners and admins can add or change Slack steps.
The model picker shows a usage limit chip next to the provider name when a spent plan window limits all its models, with the time when the limit resets. Screen readers also read the limit on each model that it stops. The picker uses the readings of the Usage rings and asks the provider for nothing new.
Usage and the usage rail show the Kimi Code plan usage: the 5-hour window and the weekly window.
Company sign-in. A server owner can connect an OpenID Connect provider, such as Microsoft Entra ID, Okta or Google Workspace, in Members → Company sign-in. After a test sign-in passes, people with an email at the company domains sign in through the company page: in the browser from the desktop app, and in the same tab in the web app. A person who signs in this way joins the server as a member with no invitation, while the server has a free seat. A person whom the owner removed stays out.
For Microsoft Entra ID, the issuer field accepts the tenant ID URL, the tenant domain, or the metadata URL from the portal.
User sync (SCIM). When the company provider turns off a person, OpenBot signs them out and removes them from the server. When the provider turns the person on again, they come back.
You can set the speed of replies that are read aloud: 0.75×, 1×, 1.25×, 1.5× or 2×. Set it in Settings → Voice → Spoken replies, with the speed button in the voice conversation panel, or with the speed button on a spoken reply. The voice keeps its pitch, and a change applies to the reply that plays now.

Improved

OpenBot downloads OpenCode 2.2.5. Before, it downloaded OpenCode 2.1.1.
Chat messages use more of a wide window. A reply, code and a table can be up to 1200 px wide, your message up to 800 px, and the chat column grows to 1920 px. Before, the limits were 720 px, 420 px and 1440 px. On a laptop the chat stays almost the same. (#4608)
A working agent in the sidebar shows one loader: the typing dots on its avatar. The face no longer shows its own felt dots. It changes to a random face at random times.
When an agent cannot attach a file to its answer, the error now gives the reason, such as a missing file, a symbolic link or a file outside the permitted folders. Before, each failure showed the same message that the file must be in the agent workspace or the shared folder.
The GitHub star request has less space under and beside its buttons, so the buttons sit nearer to the corner.
A channel summarizes long history in the background after a turn, so the next turn starts sooner.
An agent shows a notice when two compactions in a row leave its context almost full, and then stops compacting automatically.
An agent shows a notice when its provider retries a request 3 times. When the context is half full or more, the agent compacts it before the next turn (#4374).
A channel turn after the first sends less: no fixed rules again, and the channel facts only when they change (#3380).
A long request is in a channel turn once. A long handoff of an agent and a long message in channel_history are cut, and the agent can read the rest. A channel_history page also cuts a long task instruction or expected result.
An agent compacts its context before a turn whose input would fill it.
A channel summary uses the provider's smaller model when the lead uses a larger one.
Phone notifications now stay quiet while you use the computer that runs the OpenBot desktop app, also in another app. Before, the phone stayed quiet only for 30 seconds after you sent, answered or typed in the desktop app of a joined server. The phone notifies again after 30 seconds with no keyboard or mouse input.
An answered form that asked for a secret, such as an API key or a password, shows its answer as a list of rows. Each value has its label, and each secret shows as hidden. Before, the form showed empty, disabled boxes and repeated the values in a line under them.
In the model picker, Effort is a dropdown on one line, at the bottom of each provider tab. Choose an effort and then a model on that tab: the model starts with that effort.
Each file that you open from a chat gets its own tab in the side panel, after the Browser and the other panes. You can keep many files open, go between them and the browser, and drag a file tab to a new place among the file tabs. Before, a file replaced the side panel, and a new file replaced the open file.
The faces that you can choose for a new agent now have open eyes, and their eyes are larger, so you can compare the faces more easily. Before, some faces looked sleepy or angry.
The price section of the home page now says that OpenBot is free for personal use. It lists what you get, and has a Download button and a "License for work" button that opens the pricing page.
The pricing page now starts with a "Personal use is free" strip and a Download button, above the order form for servers and licenses.
The first question of first run now says that OpenBot is free for personal use. The two answers are "Use free models" and "Use my AI subscription", so that "Start free" does not look like a trial of a paid app.
The answers to "Is OpenBot free?" on the home page and the download pages now say that it is free for personal and other noncommercial use, and that commercial use needs a license.
The "Help make OpenBot better" box of first run has a shorter text. Settings still has the full text.
A full-window pile of plushes shows while OpenBot starts and while a remote server loads for the first time, on desktop and on the web. When the app is ready, the plushes hop out and the screen fades. The bar at the top of a remote server is gone: after 10 seconds, a message with Retry and Details shows.
A long draft in a confirm card no longer scrolls inside its box. The box shows the start of the draft and fades out, and "Show all" opens the rest in the card. Now only the chat scrolls.
A text paste of more than 4,000 characters in the chat box becomes a pasted-text.txt attachment at the cursor, as on the phone. The message keeps all the text, also above the message length limit. When the file cannot attach, the error shows and the text goes into the message.
The "Author is not a bot" filter of a GitHub workflow trigger now also refuses OpenBot's own bot. It checks the author of the issue, pull request, comment or review, and the account that did the action. When one of them is a bot, the workflow does not start. For example, a person who merges a Dependabot pull request no longer starts a workflow with this filter. Turn off the filter to start on these events.

Fixed

Kimi sign-in works with the newer system CLI. Before, it could fail before the sign-in code appeared. If the login process stops, OpenBot shows its error instead of an interruption error.
On a Mac, the Dock icon stays after you close the window, so its dot shows that OpenBot still runs. The menu bar icon stays too.
Kimi sign-in now opens kimi.ai, for global accounts. To sign in with a mainland China account, select "Sign in with kimi.com (China)" in the ⋯ menu of the Kimi row.
An agent that starts a task each routine run no longer stops at 8 open tasks. Only tasks that run or have a queued message count toward the limit. When the agent starts a task, OpenBot archives its finished tasks other than the newest 8. Tasks that you started stay open.
Sign-in codes reach users in all regions. Before, every sign-in code failed with an email delivery error for users whom Cloudflare served from some locations, such as FSD.
An agent with backup models and "Ask before switching" off now moves to its next backup model when its ChatGPT, Claude, Grok or Muse credit balance or quota is spent before the agent starts to write. Before, the message failed under the usage notice, and the agent stayed on the spent model.
The "New server on this computer" dialog now says that a server name needs at least 6 characters. Before, it turned off Create server for a shorter name and did not say why.
When Cursor's server stopped accepting the sign-in of its running process, every Cursor agent answered only "Please sign in to continue" until you restarted OpenBot, also when Cursor was still signed in. Now the turn fails with a clear message, OpenBot starts Cursor again, and your next message works. If Cursor is really signed out, Cursor asks you to sign in.
When Google stops accepting the Gemini sign-in while OpenBot runs, a new Gemini chat now shows at once that you must sign in with Google again. Before, the chat waited five minutes and then failed, and the saved Gemini sign-in was deleted.
The main log (logs/main/main.log) now records why a message did not start, with the redacted provider error. Before, a channel showed only "The provider could not start this assignment", and no log file kept the reason.
On Windows and macOS, OpenBot now shows a warning when another program, such as an agent, changes its program file app.asar while it runs. Before, OpenBot did not open again after a restart and showed no message. Install OpenBot again to repair it; your local data stays.
The message box shows a New chat button for an agent that has no context ring yet. Before, New chat was only in the context ring, and the ring showed only after the agent finished a turn since the app started.
A question card of an agent shows a Markdown link in the question as a link that you can open. Before, it showed the link text and the address as plain text.
When a spent Gemini quota moves an agent to its backup model, the agent now stays on the backup until the reset that Google names. Before, the agent went back to Gemini after 10 minutes, and its next turn could fail while it was working, with no switch to the backup.
In a direct chat, you can open a file that another agent names in its message. OpenBot opens the file from the workspace of the agent that wrote the message. Before, the path was only text.
An agent can now delete a webhook routine when you ask it in the chat. Before, the agent said that the routine no longer existed. To test a webhook routine, the agent tells you to use Routines settings.
The "The generated profile is too large" error no longer shows in channels and chats. Before, the long reasoning of an OpenCode or Claude model counted as answer text when OpenBot summarized a channel or chose who answers, and the message then failed (#4558).
You can add and update a Marketplace agent on a computer whose time zone OpenBot cannot identify. The error "timezone is required." no longer shows. The routines of the agent then use UTC (#4566).
In agent settings, a long value no longer covers the row name. Before, the Polish "Powiadamiaj mnie" row showed its value on top of its name. Now a value that does not fit ends in an ellipsis.
Pi agents find the tools of your MCP servers more reliably. Before, an agent sometimes looked for them with the OpenBot tool search, found nothing, and said that it had no such tool. That search now tells the agent to load MCP tools with Pi's own tool_search.
On Windows, an agent delete works on the first try also after another provider stopped, after a Pi restart, or after OpenBot started. Before, the agent's Pi or other provider session could stay open in the agent folder, and the delete failed with "The agent data could not be removed completely" until you restarted OpenBot.
On a server that is low on memory, OpenBot now closes the idle provider sessions at once when a message waits in the queue, so the message can start sooner. Before, the message could wait up to 10 minutes for these sessions to close, also when no agent was working.
Agent avatars no longer show only their eyes while the app or the website opens. An avatar now shows when its body is ready.
When Claude cannot reach its server, the chat now shows that Claude tries again, with the retry number, for example "Claude cannot connect to its server. Retry 3 of 10." Before, the chat stayed empty for minutes, as if Claude did not answer. When the connection comes back, the chat says so and the reply continues.
OpenBot loads the main window again when its page stops, for example after a crash or when the computer runs out of memory. Before, the window stayed white until you restarted OpenBot, and Ctrl+R did nothing on Windows. Agents keep running while the window loads again.
While an agent works, the dots above its face no longer cover the bottom of the message above. Before, the dots drew over the edge of that message.
A channel member always gets the current assignment and the request messages of its task. The assignment is first in each turn, and a short assignment comes again at the end (#3557, #3630).
A channel keeps its summary when a message is only marked as superseded. Before, the next turn summarized the whole history again (#3761).
A channel turn no longer sends a member its own replies or the task instruction again. The newest messages fill a quarter of the context, and older messages are in the channel summary (#3905).
A long channel no longer stops a turn while the oldest message that is not summarized still arrives. The turn gets the newest messages, and the summary goes on when that message is complete.
A channel summary uses the model of another member when the lead's model is not available (#3957).
On Windows, Connect for Gemini opens the Google sign-in page again. Before, it stayed on "Connecting" with no page and no error.
The thinking of a Claude agent keeps its words whole. Before, a line break split the words at each streamed part of the thinking.
An agent that gives channel work with a missing or incorrect value, for example no source messages or a recipient that is not a channel member, now gets the reason and can correct the call. Before, a Codex agent saw only "dynamic tool request failed" and tried the same call again, and the channel showed a "Provider error".
When many providers or accounts are connected, the provider list in the model picker scrolls. The add button below it stays visible, so you can open the provider settings from the picker.
Adding a skill with the name of one of OpenBot's own skills, such as openbot-data or openbot-skill-creator, failed on every agent with "did not change" and "An unmanaged skill already exists". OpenBot already gives every agent these skills. The local skill library now refuses a skill with such a name and no longer lists one that it took before. If an install still meets one, it gives a clear reason.
A branch thread shows "Branched from …" also when the last copied message is not shown as a row, such as a thought of the turn. Before, the line was missing in that case.
Interactive cards show in Polish when the interface language is Polish. Before, labels such as "Show all" and "Show less" on a confirm card stayed in English.
You can add a team account and a personal account that use the same email, for Claude and for Codex. Before, the second account failed with "This Claude account is already connected" or "This Codex account is already connected".
After an agent submits a form in the browser, such as a booking, it reads the page again before it reports the result. Before, an agent could say that a booking failed when the page showed that it was done, because a separate list did not update yet.
On the web, a signed-in user no longer sees the sign-in form for a moment when the page loads.
The desktop app now signs out when the account service ends its session, for example after a revocation. Before, it stayed signed in until the next start.
The app no longer stops responding while a chat history loads, on desktop and on the web.

Nov 2, 2026

Yaji 0.59.0

New

After your third message to an agent, OpenBot asks one time for a star on GitHub. The request stays until you close it, and Star on GitHub opens the OpenBot repository in your browser.
You can now attach SVG images to a chat, on desktop and in the web client. The chat shows the image, and the agent gets the SVG file and reads its text. An agent can also send an SVG back with its answer. A remote server must have this version or later to accept SVG files.
Kimi is a new provider. OpenBot downloads the Kimi Code CLI from More providers, and you sign in with your Kimi Code account in the browser or with a device code. Kimi agents use the Kimi models of your account. When a Full access Kimi agent changes its model, the Kimi CLI also saves it as its default model.
Hermes Agent from Nous Research is a provider. OpenBot uses the hermes that you installed, with your own ~/.hermes memories, skills and sign-ins. Sign in with a Nous Portal plan, or set up other providers with hermes setup. The model menu groups the models by the Hermes provider.
Add a guide to installing OpenBot on a VPS with Ubuntu or Docker, including account setup, updates and backups.
Send license emails from licenses@openbot.run. A license or a bundle gets a thank-you email at its first payment, with the license, the seats, the list price with its currency and interval, and the date of the next payment. A discounted license shows no price. A license with a free trial gets it when the first payment is made, also after a failed payment at the trial end.
Send an email when a renewal payment of a license fails, with the amount due, the date of the next try, and a link to pay the invoice. Send an email when a paid license ends, after a failed payment or a cancel. A paid renewal sends no email.
Connect a ClickUp account from the Apps marketplace to search and update tasks, docs and team comments.
A max output field for each custom model, next to the context window. OpenCode and Pi send this limit to the endpoint. Leave it blank to use the default.
Connect more than one Notion account. Use Add account on the Notion page in Apps, give each account a name, and tag the account with @ in a message. A joined server must run this version.
Connect a Link account in Server settings > Vault > Cards. An agent can then ask to pay with Link: you approve each payment in Link, and the agent fills a one-time card at the checkout without seeing its number.
Keep pinned agents and channels the same on the desktop app, the web client and the phone. The server keeps the pins of each user, and a pin, unpin or new order on one device shows on the others. The first device that connects after the update gives its pins to the server. Other devices then show the pins of the server, and their earlier pins are not kept. A server with an earlier version keeps the pins of each device, as before.
An agent can ask you to add a login, a card or an address to the Vault. The chat shows a card that opens the Vault dialog with the fields that the agent knows already filled in. You type the password or the card number; the agent gets only the name of the saved item.
Sidebar sections are now folders, and a folder can hold folders, three levels in all. Drag a chat or a folder into a folder, or use "New folder inside", "Move to", "Move up" and "Move down" in the folder menu. A nested folder shows its path, such as "Clients / Acme", and folder headers stay at the top when you scroll. A joined server and the web client need a host with this release. An app from an older release shows each folder at the top level, and its changes keep the folders.
Start a workflow when someone mentions OpenBot in a Discord server, with filters for the channel, the role and the text.
Send a workflow result to Discord: reply to the message, post in a channel, or add a reaction.
On a Mac, OpenBot now has an icon in the menu bar. After you close the window, the Dock icon hides and active tasks keep running. To keep the Dock icon, turn off Settings > General > Show OpenBot in the menu bar.
Choose the None, Minimal and Ultra reasoning levels when a model offers them, such as Ultra on the newest Codex models and None or Minimal on OpenCode and Pi models. A remote client on an older version shows the nearest level that it knows: Low for None and Minimal, Max for Ultra.

Improved

An agent now needs less supervision to finish a request. When its plan still has open steps, it can continue by itself up to 8 times, not 3. When an answer only asks "Shall I continue?", the agent continues and does not wait for you. When a turn has no plan and ends on a step that the agent named but did not do, such as "I'll now run the tests.", the agent continues up to 2 times. A question about a named action, such as "Should I proceed with the deploy?", still waits for your answer.
The model picker shows the settings that Cursor sets for each model, such as "256k · High · Fast", and shows the effort beside the model name. When an agent asks for a Cursor model value that Cursor does not list, the error names the values that Cursor lists for that model.
Global search (⌘K) now lists agents in the order of the sidebar: pinned agents first, then each section from top to bottom, then hidden agents. When you drag or pin an agent in the sidebar, its position in ⌘K and its ⌘1-9 number change too. An agent in "Needs you" keeps the place of its section, so the numbers do not move when an agent starts or stops waiting (#4430).
The Version line of a provider in Settings says if the CLI is the one that you installed or the one that OpenBot downloaded, also when that CLI fails.
A pasted image shows as a thumbnail above the text of the message box, and its name stays in the text where you pasted it. This is so in agent chats, channels, side threads, side threads beside the chat, and on the Tasks page.
An image that you send shows full size above your message, outside the bubble. A message of images only has no bubble. Before, the image showed as a small file card or only as its name.
A pasted image that takes time to add shows a preview with a spinner in the thumbnail row until it is ready.
You can paste images and files when you start a task on the Tasks page, and in the message box of a task.
In the web client, an attached image shows its preview in the message box, and images in messages show full size. Before, they showed as files.
Side threads take files on a joined server and in the web client when the host's OpenBot can take them. Before, only this computer took files in a side thread.
The switch that pauses a routine in the routine list is larger.
Basic product analytics and error reports, such as turns and token counts, now go also when analytics is off. They are anonymous: they carry a random install ID and no name, email or account. The ID changes after 13 months. Chats, files and prompts are never sent. The setting is now called "Link analytics to my account". The websites that you and your agents visit, tool use and a daily summary of your setup go only when it is on.
On a Mac, one press of ⌘Q no longer quits OpenBot. Hold ⌘Q or press it twice to quit, and a hint in the window tells how. Before, one press stopped OpenBot with its agents and routines, and it was easy to press by mistake. Quit in the OpenBot menu, the Dock and an update still quit at once.
When a process stops OpenBot with a signal, such as kill, OpenBot now stops its agents and saves its state before it quits. Before, the desktop app stopped at once.
The main log now says how OpenBot quit: from the menu, with ⌘Q, from the tray, for an update or by a signal.
Agents that have no photo now show a soft plush bot instead of the flat drawing, in the app, the web client and on openbot.run. When you choose a new face or colour, the old plush shrinks away and the new one pops out.
The plush shows what an agent does on its face, with no coloured rings or glow round it. A working agent looks up while three felt dots pulse, an agent that waits for you shows a felt dot, a failed turn lowers its eyes and shakes "no" once, and an asleep agent closes its eyes as z's rise. When you point at a resting agent or move the keyboard focus to it, it shows a different friendly face each time (a wink, wide eyes or a curious look) until you move away.
In agent settings you now choose the shape and the eyes of an agent's face separately. The shuffle button picks a new pair. If the colour is Automatic, new eyes keep the colour that shows, and that colour becomes the chosen one.
The avatar colour list shows each colour one time, in one row.
An agent photo in the desktop app shows as it does on mobile: a plain circle, with no ring or added background.
The loading screen now shows felt plushes. Each time, a different crew falls into a pile and the OpenBot plush lands on top. You can grab a plush and throw it. When the app is ready, the crew hop out of the screen.
While the history of an agent loads, the OpenBot plush watches three felt balls bounce. Billing, hosted servers and the GitHub and 1Password connector steps show the same plush while they wait.

Fixed

A cloud server that installed an update now starts after a later wake. Before, the server could show "Starting Cloud server" for many minutes and not connect, because the update could leave OpenBot files that the server could not run.
The Usage page shows the daily cost and token chart again. Before, the area next to the total was empty.
Pi agents now show the context meter and compact their chat at 80% of the model's context window, as Claude and Codex agents do. Before, a Pi chat grew until Pi's own limit, because OpenBot did not read the context fill that Pi reports (#4413).
Slack mentions and direct messages reach OpenBot again. Before, when a computer with a Slack workspace was off, each channel message of that workspace failed in Slack. Slack then stopped sending events to the OpenBot app for all workspaces, with no error in OpenBot.
Codex agents get the tools of an MCP server that starts slowly, for example an npx server after OpenBot starts. Before, Codex waited 1 second for the server, and the first turn of the session did not have its tools.
When an agent asks for a saved password or a code on a sign-in form inside a frame of the same site, it now gets the reason: secure input does not support frames, so it uses takeover. Before, it was told that the field was not an editable input, and it tried other fields.
A file that uploads to a chat no longer blocks Send in the chats of other agents. Before, a large upload to a server stopped Send in each chat until it was done.
In the web app, a file that you add while another file uploads now shows "Wait for the current attachment to finish." Before, the file waited and could go to the wrong chat.
An agent with backup models and "Ask before switching" off now moves to its next backup model when its Gemini quota is spent before the agent starts to write. Before, the message failed with the Gemini usage-limit error, and the agent stayed on the spent Gemini model.
After the computer wakes from sleep, phones and other devices can connect to its servers again at once. Before, the server could stay unreachable for up to 45 minutes, until OpenBot restarted.
A new custom endpoint now shows its models in the model picker when no agent uses OpenCode or Pi. Before, OpenBot stopped the idle provider and did not start it again after the save, so the picker showed the endpoint with no models until OpenBot restarted.
A provider CLI update could fail with "OpenBot could not activate the managed CLI." when you also installed the CLI yourself, as for OpenCode. The update checked your CLI in place of the new version, and showed its version or its failure as a failure of the update. The update now checks the new version. When the new version does not start, the error says so.
In the browser client, a server that refuses an action, such as an ACP registry install, now shows its reason. Before, the browser showed only "The server could not complete this request".
Keep the focus in the chat box while an agent opens pages in the built-in browser. Before, each new page took the focus, so the text you typed stopped or went into the page.
Text that you copy from OpenBot now pastes into mail, documents and chat apps without the styles of the app. This applies to Cmd+C or Ctrl+C in a chat, in the message box and on a formula, to a cut in the message box, and to the Copy button of a message with attachments. Headings, lists, tables, links, code and spaces stay, and line numbers and button labels do not come with the text. Before, the paste had the colors of the app theme, such as white text on a dark background, and you had to paste without formatting.
An agent on an OpenCode model with the z.ai GLM Coding Plan now moves to its backup model, or waits for the window, when z.ai says "Usage limit reached for 5 hour" or "Weekly/Monthly Limit Exhausted". Before, the turn failed as a normal error.
When a Workspace only custom agent fails because the macOS sandbox blocks a write, for example to its own database in the home folder, the error now says "Workspace only blocks this. Try Full access." Before, only the agent's own error showed, with no cause.
In the routine list of an agent, hover on a routine now shows one colour on its card. Before, the routine name and the switch area showed two different colours.
A routine stays in its position in the list when you pause, resume or edit it. Before, it moved to the top of the list.
The live view of a remote host's browser tab now shows the full page and puts each click where you point. Before, on a page larger than 1280 by 800, the picture filled only the top left of the view and the pointer moved away from the target towards the bottom right.
The live view of a remote browser tab now stays close to the pointer on a busy page. Before, the host sent every pointer move to the page, so the page could fall a second or more behind, and a click waited for all of the moves before it.
A custom model with an output limit below 32,000 tokens can run a turn on OpenCode again. Before, OpenCode asked the endpoint for 32,000 output tokens on every turn, and the endpoint refused the request. A model with a context window and no output limit now gets at most 16,384 output tokens, as on Pi.
Typing and click sounds no longer play as one very loud burst. Before, when the audio output paused for a moment, the sounds of all the keys that you typed in that time could play at the same instant.
You can delete an agent that a usage limit holds. Before, a message that waited for the limit to reset blocked the delete with "Stop the agent and cancel its queued messages before deleting it." The waiting messages are now deleted with the agent.
A hosted or self-hosted server on a Linux kernel older than 6.1, for example in a Proxmox VE 7 container, can open network connections again after an update. Before, the AppArmor profile that the update loads blocked each socket (#4335).
The status line under the agent now translates in every language. Before, the generic tool steps ("Working through the next tool-assisted step…", "Reviewing the latest tool result…") stayed in English in a Chinese, Japanese, German or any other non-English UI.
A server (self-hosted, Docker or hosted) now uses less CPU when it is idle. Before, OpenBot drew its window on the virtual display of the server, where no one sees it, and on a server with few cores this used a large part of one core.
Saved logins from the Vault, 1Password, and Bitwarden now fill the password on sign-in pages whose form has no method, such as Clerk sign-in pages. Before, the password card opened each time. OpenBot still stops the browser from putting a filled password in a page address.
When OpenBot does not fill a saved login, the agent now gets the reason, and the log shows it. Before, the password card opened with no reason.
Messages that wait in an agent's queue now stay in the queue when its provider must be signed in again, and they start after the sign-in. Before, a provider check that found the provider signed out during a turn marked those messages as failed, and they were never sent (#4326).
A failed channel routing turn now writes its redacted cause, with the lead's provider and model, to main.log as channel_coordination_failed. Before, this line was missing when the channel showed "The channel lead could not route this message" or "The channel lead did not answer in time".
On macOS, the Dynamic Island no longer makes OpenBot the active app when you switch to an app that has no window on the current Space. Before, OpenBot came to the front about half a second later.
When a turn ends or stops, or a provider stops, while an approval card is open, OpenBot now answers the provider that asked. Before, the card disappeared, the provider waited for an answer that never came, and the agent could hang. An agent that uses ACP, such as OpenCode, now gets a cancelled answer, not a refusal.
A mention of the OpenBot role in a Discord server now reaches OpenBot, as a mention of OpenBot does. Discord's suggestions often choose this role instead of OpenBot.
The file hint in the avatar editor stays inside the upload card.
When you change an agent's avatar, the open conversation no longer draws its agent tags again, so the change is smoother.
Agent faces sit level with the text next to them in the "working" line, the waiting line, the sidebar, search, agent tags in messages and the direct conversation header.

Nov 1, 2026

Yaji 0.58.0

New

Show the OpenCode Go monthly limit as a third bar in Usage, under the 5-hour and weekly bars. Before, Usage showed only the one of the weekly and monthly limits that was closer to its end.
Add 17 OpenBot agents to Marketplace: QA Tester, PR Shepherd, Skill Smith, Prototype Builder, Design Critic, Accessibility Desk, Handoff Spec Writer, Copy Editor, Search Content Planner, Prospect Researcher, Call Coach, Chief of Staff, Spend Auditor, Request Triage Desk, Ops Digest, Home Care Log and Bill Tracker. Each agent does one job and asks before it sends, spends, merges or publishes.
Add 15 OpenBot skills for these agents, such as Acceptance Check, PR Follow-up, Voice Edit, Daily Brief, Spend Audit and Care Log.
Some agents install with an active routine, such as the weekday morning brief of Chief of Staff. A routine asks once for its setup and sends nothing when there is nothing new.
Start a workflow from GitHub. A GitHub trigger runs on new issues, comments, labels, pull requests, reviews, merges, closes, assignments, failed checks, or a mention of @openbotgit, in a repository that the OpenBot GitHub App can see. A second event for the same issue waits for the first run, and events for other issues run at the same time. The trigger shows the newest 50 events and why each one started a run, waited, or was skipped.
Change GitHub from a workflow. A GitHub output comments, adds or removes labels, assigns a person, asks for a review, approves or requests changes, closes, reopens, or creates an issue, as the OpenBot GitHub App. An agent can choose the labels or the person from the repository's own lists.
Show each open run on the workflow canvas, and the events that wait for one.

Improved

The server invitation page has one Open OpenBot button. Below it, links let you join in the browser or download OpenBot. When OpenBot does not open on a computer, a dialog tells you to install it and gives a download button.
The invitation, Slack and Discord connect, billing and report pages no longer show a blue glow and a shadow behind the card.
Rewrite the OpenBot agents and skills in Marketplace. Each one now has a named result, the evidence it gives for each claim, and a fixed rubric. Agents that you installed get the update offer in Marketplace.
Remote connections now use two Signal machines at the same time. Each server has a home machine. When one machine stops, its servers and their devices move to the other machine in about one minute.
The first conversation of a joined server opens faster when the server has many agents. Before, the app showed it only after it read the queue of each agent. On a server with 50 agents, the desktop app now shows it in about half the time.
The web app reads the agent settings, connectors and routine calendar of a server once when it opens, not up to three times. On a server with 50 agents, its side panels show about a quarter sooner.
When you open the app again, the agents and sidebar of a joined server show at once from the last visit, while the app connects to the server. The saved view holds no messages. Server actions start when the connection is made. The channels show in their place too, and a hidden or moved agent stays hidden or moved in the saved view.
The desktop app opens its window about 3 seconds sooner when the Dynamic Island is on. The window does not wait for the Island to load.
A client connects to a joined server with fewer steps in sequence. The desktop app reads the server compatibility once, not twice. The web app loads its workspace code while it signs in, and on a server it connected to before, it reads the workspace at the same time as the compatibility. With 100 ms network latency, the web workspace is ready about 250 ms sooner.
On the web, a switch to another server is about 2.5 times faster. The app uses the connection that it already keeps to that server for notifications, and it does not wait for the old server's session to end. On openbot.run, a switch took 0.7 to 1.0 seconds, against 1.7 to 2.8 seconds before.

Removed

Remove six OpenBot agents from Marketplace: Product Design Lead, Communications Partner, Meeting Assistant, Account Research Partner, Outreach Writer and Operations Planner. Other agents now do their jobs. Copies that you installed keep working.

Fixed

A custom provider that you add now appears in the model picker when OpenCode or Pi is not installed. OpenBot downloads the provider that runs the endpoint when you save it. Before, the endpoint was saved, but the Custom tab stayed hidden until you downloaded OpenCode yourself.
The account service does less database work when a phone or browser connects to a server. Before, each connection read the connections of all users of that server, and each server or account deletion read every connection of every server.
Accept the Claude sign-in check when Claude reports that it is signed in but its status text is not readable. Before, this stopped the turn of an agent with its own Claude account, and the message to that agent failed.
The download button in the "Nothing opened?" dialog of the invitation, Slack and Discord pages now uses your language. Before, it was always in English.
Moving a queued message works when more than 32 messages wait in the queue, on this computer and on a joined server. Before, every move failed with "Invalid delivery ids." or a 400 error and the order did not change.
A Pi agent turn now fails with a message when Pi sends nothing for 10 minutes. Before, the turn stayed "working" until you stopped it, and other messages waited behind it.
Long model names in the model picker, such as the Fusion models of Devin, now go to a second line. Before, the picker cut off the end of the name.
In the web app, a server that is open in another browser tab now shows a full-window screen that says so, with "Use here" to move the server to this tab. Before, the page showed "Starting" for 15 minutes, started the server again every few seconds, and then showed "did not start in time", while the server ran.
A host that starts lists its agents as soon as they load. Before, a client could wait about 15 seconds for the list while the host ran its routine checks.
On the web, an agent that you open while the app connects stays open after the connection.
While a joined server connects, the conversation says that it connects. It does not ask you to complete the provider setup.
While a joined server connects, a channel from the last visit does not say "No messages yet" or show its creation date.

Oct 31, 2026

Yaji 0.57.0

New

An agent can now answer a website's alert or confirmation that its own browser action opened. Before, the dialog stopped the agent until you answered it. A dialog that you open, or one that is still open when the agent's turn ends, stays for you to answer.
An owner or admin of a joined server can now turn Agent starts new chats on or off for an agent from the mobile app. The computer that runs the agent must have this version. A change on the phone shows at once in the agent settings on the computer, and a change on the computer shows at once on the phone.
An owner or admin of a joined server can now turn "Show the plan before long tasks" on or off for the server's agents, in the agent settings on desktop and in the web client. Before, only the computer that runs the agent could change it. The server must run this version.
A phone that runs the new mobile app now gets the time at which the plan card starts by itself, so it can show the countdown and say when nobody answered in time.
An owner or admin can change the folders of an agent from the mobile app: add, remove, exclude and include again. The host checks each folder as the desktop folder picker does. A change on the phone or on the desktop shows on the other device at once. The computer that runs the agents needs this version.
The owner of a server can change its server key from the mobile app. The server on your computer makes the new key and signs the change with the current key and the new key, so paired devices stay paired. Update OpenBot on the server computer to use it.
Server settings → General → Server key shows when the server last changed its key. Every member sees it. Only a change that the previous key and the new key both signed shows.
The phone connection can now receive a transfer file of up to 100 MB from a server, so the mobile app can export a server. Other files that the app downloads keep the 10 MB limit.
Agent settings have an Instructions page that shows the agent's AGENTS.md file as formatted text, with the file name and its folder at the top. Edit opens the whole file on the page, and Show in Finder shows the file. The agent keeps its standing rules in this file, as short bullets under headings, and changes the file when you ask it to change its instructions. A workspace that has only a CLAUDE.md file keeps that file. Each provider gets the file, also a provider that does not read it by itself.
When the agent changes its instructions, the chat shows "Updated its instructions" after its answer. View opens the instructions, and Undo puts back the text from before the change.
You can branch a conversation with an agent. On the last agent message of a turn, select the branch button, or right-click the message and select Branch from here. A new thread of the same agent opens with the conversation up to that message, and its message box is ready for you to type a different direction. The original thread does not change. A line at the top of the new messages shows the thread that you branched from; select it to go back to that message. Files in the agent's workspace are not copied or rolled back. Joined servers and the web app can branch when the server has this version.
An owner or admin can add, change and remove the custom ACP agents of a server from the mobile app. A server that runs this version accepts these changes. Each agent's command runs on the server's computer, and the values of its environment variables go only to the server.
A stopped or failed channel task now has a Cancel task button next to Continue and Reassign. It ends the whole run of that task, and keeps the messages. Before, a stopped task stayed above the composer until you continued or reassigned it.
Disconnect Grok, Muse and Cursor from Settings > Providers, in the ⋯ menu of their sign-in, as Claude and Codex already do. On this computer, the CLI in your terminal also signs out. A joined server must run this version to disconnect them.
The Marketplace is now on openbot.run at /marketplace, with the Agents, Skills and Apps tabs. Each agent and each skill has its own page. A shared link shows a picture of that listing.
Agent and skill pages in the app have Copy link. The link opens the same listing on openbot.run, and its button opens it in the app. A link never adds the agent or the skill: you still press Add.
The Agents tab in the app shows Featured agents, and a switch between Verified and Public agents.

Improved

Settings > Notifications shows one card for each server, with All, Needs me and Nothing as one control. Only the agents that have their own level show under their server. "Choose another agent" gives one more agent its own level, and the X button gives it the server level again.
An agent that a member turned off for everyone in an earlier version turns on again when you give it its own level. Settings no longer shows "Off for everyone on this server".
Memories now open in the settings pane of an agent or a channel, not in a dialog. The list shows a short title for each memory, in "Always used" and "Used when relevant" groups. A title opens the memory on its own page, where you read, edit or delete it, and where a "Use in every chat" switch sets whether the agent always uses it. A search field above the list finds memories by their words.
Agents now start a long memory with a short label, such as "Release notes:", so the list shows a clear title.
Show a browser profile change made on another device at once in the desktop and web Browser profiles settings and in the agent settings. A host that runs this version tells each client about each change.
The old Instructions field is now Description: a short summary that teammates read to choose who does a task.
When a tab or the thread menu cuts a long thread title, a number at its end, such as #2, stays visible.
The reaction button below a message now has the same hover color as the copy and reply buttons.
Remote connections that need a relay server now spread over two relay servers. When one relay server is full or stops, new connections use the other one.
Remote access now uses two connection servers at the same time. Each server and its devices meet on one of them. When that connection server stops, the server and its devices move to the other one automatically.
Show the OpenBot server offer at the bottom of the Routines panel, below the routine list, with its buttons on the right. Before, it was above the routines.
A chat shows one schedule card for each routine: the newest change by the agent. Earlier changes show as one line that opens the routine.
Show each routine in the agent's Routines list as its own card, with a switch at the end that pauses or resumes it. The Pause and Resume buttons are gone.
When you pay for a plan or a license, the account service sets your account email on your analytics profile, also when product analytics is off on your devices. See PRIVACY.md.
Send an attached MP3 file to an agent as audio when its provider accepts audio input, such as Antigravity. The agent hears the file and still gets its path. Audio over about 14 MB a message goes as a path only, as before.
Put more space between the top bar and the title of each Marketplace list and page.
Send about 3k fewer tokens in each Claude agent request. Claude agents no longer get the Workflow and ListAgents tools, which OpenBot does not show, and Claude auto memory is off: OpenBot memory is the agent's only memory. Earlier Claude memory files stay on disk.
Send fewer tokens in each request of the other providers: - Pi loads the tools of your own MCP servers only when the agent searches for them. With computer use on, this removes about 24k tokens from each request. - Grok no longer gets its send_feedback and workflow tools. - Antigravity no longer gets its schedule and run_workflow tools. Use OpenBot routines. - Codex goals are off. OpenBot has no control to pause or stop a goal. - OpenCode no longer gets its tools that rename or move a session, or that make or list worktrees.
The agent page in the app shows the skills, routines and information of the agent in sections.
openbot.run/plugins now opens the Apps tab of the Marketplace.
When an agent's browser action matches more than one element, the browser still sends no input, and the error now lists up to five of the matches with their text and the text of their item. The agent can then act on the intended item, for example one lesson in a calendar, or ask you.

Fixed

Update the file preview panel when the agent changes the open file or folder. Before, the panel showed the file as it was when you opened it.
While the phone downloads a file from a server, other answers from that server no longer wait for the download to finish.
An agent with no title no longer saves "General assistant" as its title.
A computer that joined a server over a direct connection now gets the newer server features, such as thread branches and notification levels for agents. Before, it got only the older features of the server.
An agent's change to a routine schedule stays saved. Before, when the agent changed a routine to a schedule that only the cron field can show, each older routine card in the chat saved the earlier schedule again, and the chat filled with "Updated routine" lines.
Install an in-app update on a slow Mac, such as an Intel Mac. Before, OpenBot closed 30 seconds after "Restart to update", before macOS finished preparing the update, so the old version opened again and offered the same update.
An agent on an OpenCode Go model now moves to its backup model when the Go usage limit is reached, and continues the unfinished work there. Before, the turn failed with "Go usage limit exceeded", and each retry went to the same model again.
Change the language again from the sign-in screen. The button under the form shows the current language and opens the language choice. Before, a language chosen by mistake stayed until sign-in.
Fill the macOS installer window with its background. Before, a white strip showed below the background when Finder showed no tab bar or status bar.
In the sign-in dialog for a provider on a joined server, such as ChatGPT on a cloud server, the "Stored on" line no longer touches the "Open sign-in page" button.
OpenCode now starts again by itself when a start fails with "Internal service failure", as it can on a slow computer. Before, OpenCode stayed off and messages waited until you refreshed the providers.
Deleting a side thread while Codex compacts its context no longer holds the agent's own chat. Before, every later message to the agent stayed queued until OpenBot restarted.
ChatGPT no longer shows a "Codex configuration warning" at each start when the shared Codex configuration has a setting that Codex does not know, such as the Windows app approvals that the Codex desktop app saves. Codex works without these settings. Before, the warning told you to remove the setting, which made the desktop app forget its approvals. The warning is now in the log.
ChatGPT no longer shows Codex log lines as "ChatGPT error" messages while it works. Before, a line such as the notice about a folder that is not trusted showed again at each Codex start. When Codex stops with an error, the message now shows its last error line.
Choose a server's custom provider models from the Custom tab in the desktop app. Before, the model picker of a joined server's agent used the endpoints of this computer, so the Custom tab was missing and the server's custom models showed on the OpenCode tab.
When a custom endpoint refuses its API key, or needs a key and has none, the error now names the endpoint and tells you to enter the key in Server settings → Providers. Before, it said that OpenCode had no sign-in and asked for an OpenCode Go key.
An agent's browser tab that is idle for 30 minutes now ends its renderer process, not only its page. Saved tabs also start with no renderer process until they are used. Before, each such tab kept about 85 MB, and a server with many agent tabs used more than 1 GB for them.
When you ask an agent to connect or use an app such as Figma, Notion or Linear, the agent now shows the Connect card in the chat more reliably.
When the app is already installed and does not work, the agent now tells you what to do. It shows the Sign in card, gives the setup steps for an app on your computer such as Figma, or says that the app is turned off. Before, the card only said Connected.

Security

The web client keeps the first key that it sees for each server. A later key change is used only when the previous key signed a handover to it, as on desktop.

Oct 30, 2026

Yaji 0.56.0

New

Server settings > Storage now counts the CLI copies of providers that are off in the Provider runtimes row, with the old CLI versions. Clear removes all of them in one step. A provider that is off downloads its CLI again when you turn it on.
Right-click an agent message in a chat on this computer and select Mark unread. That message and the messages after it are unread again, the NEW line moves above it, and the agent shows an unread count in the sidebar until you open the chat again.
When you have another server, you can hide the Local server. Open its menu and select Hide server. To show it again, select Show Local in the menu where you add a server. Its agents, chats and files stay on your computer.
Agents can do many browser actions in one step, such as typing a search, pressing Enter, and waiting for results, or paging through a list until they find an item. This uses fewer tokens, because the agent does them in one call and reads the page once at the end. You can still take control of the tab at any time.
The server plans show how many hosted sites each plan has: Starter 3, Standard 10 and Pro 50.
Server settings > Sites shows See plans when a server on your computer is at its site limit.

Improved

The website now says that the source code of OpenBot is open on GitHub, not that OpenBot is open source.
A provider uses the CLI that you installed when it is the same version as OpenBot's copy or newer. Before, OpenBot's copy always ran once it was downloaded. Update still downloads a newer version for you. Server settings > Storage can then remove the copy that no agent uses.
Agents find and use OpenBot tools in fewer steps, so browser tasks finish sooner and use fewer tokens.
When the connection service restarts, each app waits a random 0.5 to 4 seconds before it connects again. Before, all apps connected again in the same half second, and many of them had to try again. Open chats and connections to your servers stay up while the app waits.
The Instructions field in agent settings and channel settings has a grip in its bottom-right corner. Drag it, or use the arrow keys, to make the field taller. Double-click the grip to go back to the start height.
Settings search finds Providers and Import, and opens them in Server settings. The command palette has a Providers entry.
First run says that you accept the Terms of Service when you continue, with links to the Terms of Service and the Privacy Policy.
First run has a "Help make OpenBot better" box. In the United States, Mexico, Japan, Australia and New Zealand it starts ticked; everywhere else it starts clear, and OpenBot sends no analytics until you tick it or turn it on in Settings. OpenBot reads the region and time zone of your computer to choose. If you use OpenBot now, your analytics setting does not change.
When you tick the box, three agents pop up to say thank you, and a toast thanks you once.
The web sign-in has the same box, directly under the send button, and the Terms of Service notice at the bottom. The web app reads the region of the browser language and the time zone to choose. Your answer is saved when you ask for a sign-in code. After you agree once, the sign-in screen does not ask again. If you are signed in to the web app now, your analytics setting does not change.
Toasts have no drop shadow.
Your analytics answer follows your account. When you tick or clear the box, or change the setting in Settings on desktop, web or mobile, the account keeps your answer, and your other devices use it after sign-in. A box that you did not change is not sent. An older choice from a device that was offline does not replace a newer one.
The icon buttons at the top of the file preview panel, such as Copy, Download and Close, now show their name in a tooltip when you point to them or move to them with the keyboard.
Choose the sound that plays when a task finishes: Off, Classic, Warm, Mechanical, Bubbly or Clicky. Classic is the sound from before. The task sound plays also when action sounds are off. In Settings > Notifications, both sound settings are now menus. Select the play icon before a sound name, or move to it with the arrow keys, to hear it before you choose it.
If you turned analytics off before this version, an analytics answer from another device does not turn it on again on that device. Only a change on that device turns it on.

Fixed

A channel lead now routes a request in a channel with many members that have long descriptions. Before, routing stopped with "This request exceeds the routing context limit", also for a short request.
A Workspace only Claude agent on a computer where Claude cannot start its sandbox, such as Windows, now says that the sandbox is not available and tells you to choose Full access or another provider. Before, it showed Claude's message about a sandbox.failIfUnavailable setting that OpenBot does not have.
In the sidebar, a chat that you drag now lands in the section that is below the pointer. Before, a drop near the edge of a long section could go to the next section, and a chat dragged into another section could go back to its own section.
OpenBot keeps one version of each provider CLI that it downloads. Before, it kept the version before the last update too, so each CLI used about two times its size on disk. The old version is removed a few hours after an update.
Cloudflare checks now pass in the built-in browser. Before, the check stayed on "Performing security verification" or asked again after you completed it.
Agents that work in a busy web page compact their conversation less often. Before, each click sent the full page details and the same page log again, and an agent could compact its context every minute.
A click, key press or other browser input now returns a short status (address, title, focus and new errors), not the full page. The element references from the last page read stay valid, so an agent reads the page again only when it needs to.
An agent can take a screenshot of one element of a web page. The image is much smaller than a screenshot of the full page.
Selecting an option in a list that the web page replaces after each change now selects the correct option. Before, the agent selected the wrong option, for example on a shop's sort list.
A page read is approximately one third smaller: each element is one line of text, and successful network requests and actions are not in it.
Browser instructions that an agent gets when it finds a tool are approximately four times shorter. Sign-in and payment rules come only with the tools for these steps.
A page read does not attach a second image of the page when the agent got one just before and nothing changed the page since.
A web request that the page cancels, for example a live price stream on a trading site, is not shown to the agent as an error.
In a channel, a turn that the provider starts by itself, as when a Claude or ACP background job ends, no longer takes the member's next task. Before, that turn ran as the next task, its text showed as the task's reply, the task's own request did not start, and the channel showed "The provider could not start this assignment."
After an agent publishes, replaces or deletes a site, the chat shows one row with the latest status. Before, a "Deploying site" row with a spinner stayed next to the "Published site" row.
Show only the providers that you connected in the model list of a workflow LLM or agent step, as the agent model list does. Before, the list showed ChatGPT and its models when ChatGPT was not connected, and a new LLM step took a ChatGPT model. A step keeps the provider that it has.
Show a neutral mark for a provider that this OpenBot does not know in a workflow step. Before, it showed the ChatGPT mark.
The Gemini sign-in hint names the correct place: Server settings > Providers.
A site that a computer published before it became a server now shows in the server's Sites list and counts against the server. Before, the list was empty and showed 0 sites. The site moves to the server the next time the server connects to your account.
An upload that stopped before it finished no longer takes a site slot, so a publish no longer fails when Sites shows a free slot.
When the site limit stops a publish, the agent and the app now say what the limit is and how to publish more sites.
A message that you send to an agent at a usage limit now tries the provider again at once. Before, it waited for the reset time, also after you signed in to another account or got more usage.

Oct 30, 2026

Yaji 0.55.0

Action needed after you upgrade

In a Telegram group that you connected before this fix, remove OpenBot from the group and connect the group again from OpenBot, or make OpenBot a group admin.

New

In Workspace only, you can now exclude folders on an agent's Folders page. The agent cannot change files in an excluded folder, also when it is in the workspace, the shared folder or an added folder. Claude and the sandboxed providers block these writes. Codex gets the exclusions as instructions only.
The Marketplace Skills tab shows "Your skills" first: each skill that your agents have, with the agents that have it, and an Update button when a new version is available.
Choose the agents of a skill in one dialog. Search 50 or more agents, select a range with a drag or with Shift+click, and save all the changes with one button.
Show only verified skills or only public skills with the All, Verified and Public switch.
Each skill group has its own page with three featured skills, a sort by use, date or name, and Load more.
A skill page lists each agent that has the skill, with a switch to turn it on or off, Update and Remove. A skill that you made on this computer has the same page.
Remove a skill from one agent or from all its agents. Undo in the notice takes the change back before the skill is removed.
While an agent works, press Tab in the message box to send the message into its running turn. Before, you sent the message and then selected Steer on its queue row.
You can dictate a message in a side thread, as in the agent's own chat. Click the microphone, speak, and the words appear in the thread's message box.
A side thread's message box has the same "+" menu as the agent's own chat: attach an image, use a skill, add context, or ask a side question.
Settings → Voice has a Voice server section. Connect an OpenAI-compatible server, such as Speaches or Kokoro-FastAPI on your network, with its address and an optional API key. Then use it for dictation, for spoken replies, or for both, with the model and voice that you choose. Nothing has to be downloaded for a part that the server does.
You can open a thread beside the agent's chat. Drag its tab from the chat header to the side panel, or right-click the tab and choose Open beside the chat. Drag it back to the chat header to open it there again. The side panel keeps its thread tabs after a restart.
You can create more servers on this computer. Each server is a separate workspace with its own agents, chats, files and settings. To add one, click the plus button on the server rail, then click "Other options" > "New server on this computer". You can have up to 5 of these servers.
Settings > Providers has Disconnect in the ⋯ menu of the Claude and Codex sign-in, on this computer and in Server settings of a server. It signs the provider out and removes the stored sign-in. On this computer, the claude or codex CLI in the terminal signs out too (#3487).
Log in to ChatGPT with an OpenAI API key. On this computer, open Settings > Providers > ChatGPT and select Add key under API key. Codex saves the key, and OpenAI bills each request to it instead of your ChatGPT plan. Change key replaces it, and Disconnect removes it.
Settings › Agents now lists the background processes that agents left running on your Mac or Linux computer, such as a dev server that an agent started with nohup. Each row shows the command, the agent, the start time and the ports, and Stop ends the process and its child processes. Before, such a process kept running after the chat ended and after you quit OpenBot, and only a terminal could find and stop it.
A custom endpoint can now run on Pi. Open Settings › Providers › Custom endpoints, edit the endpoint, and set "Runs on" to Pi. Its models then show in Pi's model list. Endpoints that you saved before keep running on OpenCode. When Pi is not downloaded, choose it in "Runs on" to download it. On a joined server, the model picker also knows which endpoints run on Pi, when the host has this version.
Settings › Notifications is now the one place for notifications: one table shows each server and its agents, with All activity, Only when it needs me, or Nothing, and a mute for each server. The level of a server and of each agent is yours only and applies on your computer, in the web client and on your phones, and a change on one device shows at once on the others. The server menu and the agent settings show the level and open this table; Server settings no longer have a Notifications section. A mute still applies to the device where you set it. The agent settings no longer have their own Notifications switch; an agent that was switched off for everyone shows "Turn on" in the table.

Improved

In a channel, more than one agent can now answer. When you ask for the view of each agent, speak to the whole group, or need the work of several agents, the lead gives the request to each of them, and each one answers in its own reply. A message that starts with more than one mention, such as "@Writer @Editor what do you think?", goes to each mentioned agent, and a message that starts with @everyone goes to every agent in the channel. The mention picker shows an "Everyone" option in channels. "continue" after a stop continues the work of each agent that the request went to. Before, only one agent answered.
The Skills tab shows 6 skills for each group, with "See all" for the full group. It replaces the install menu and the Installed filter.
Pages in the Marketplace window move with a short transition. When reduced motion is on, the page changes at once.
The Tasks page is back on desktop and the web, pinned to the page rail. It lists the side threads of all agents in Needs you, Working, Waiting and Archived, and starts a new task. A task opens beside the list in the same thread view as the chat, with its steps, Archive and Restore, and Open in chat. A task that waits for your approval or answer moves to Needs you. A task that you read there shows as read in the chat header and on the phone.
In a side thread, Stop replaces Send while the agent works and the message box is empty, as in the agent's own chat.
While an agent asks you a question, a message that you send from the chat box is your answer to that question. Before, the message went into the running turn and the question stayed open. A channel message, a message with files, or a reply to a message still goes as a message, and a private answer still uses its own field.
An agent that ends a turn with open steps in its plan now continues by itself, up to 3 times for one request. It stops when the plan does not change, when it waits or asks you, or when you stop it or send a new message. Steps that were already open before your request do not count. A short line in the chat shows each automatic continuation.
Agents now finish the whole request before they end a turn. They do not stop only to report progress or to ask whether to continue.
The thread tabs in the chat header now look and work like the side panel tabs. The agent's name is the first tab, for the agent's own chat. All thread tabs have the same width. When the tabs need more room, the model picker shows less first: without the effort, then only its icon. Then the tabs get narrower together, and after that they scroll. Thread tabs in the side panel have the same width.
In a narrow chat header, the open thread shows as one button with the look of a selected tab, and the model picker shows only its icon. When the header gets narrower again, the agent's name, then "+", then the thread list button hide, so no control covers another.
You rename a thread tab in the tab itself. The thread icon stays, and the field does not cover the next tab.
"Join a server" is now in the "Other options" menu at the top of the server plans.
The sidebar shows larger avatars: 46px in the agent, channel and people rows, and 56px in the pinned row.
Group chats show larger author faces beside messages: 32px instead of 24px.
MCP servers that OpenBot signed in to, such as Notion, now see OpenBot as the client, with the name of the agent that connects, for example "Research Bot (OpenBot)". Notion shows "with MCP" until it adds OpenBot to the clients it names.
Code blocks in chat now wrap long lines, so you can read an email draft or a long command without scrolling sideways. Line breaks and Copy do not change.
A notification with a preview, on the computer, in the web client and on the phone, now shows the first sentences of the agent's message, not all of it. It leaves out code blocks, tables, headings and mail headers such as "In-Reply-To", shows tool names such as "openbot.revise_skill" as words, and shows a hidden email address or secret as "•••" in place of "[redacted-email]". The iPhone Live Activity that the computer updates shows the same short text.

Fixed

On a joined server, a second ChatGPT account now connects after its sign-in in the browser of your computer. Before, the sign-in screen stayed open after the browser page closed, and after 10 minutes the server removed the account.
When you ask, agents run your own browser scripts, such as Selenium. Without your request, they still use the OpenBot browser.
An owner or admin of a joined server can now choose Queue or Steer for each agent of that server, in the agent's settings › Advanced › While working. Before, the row was hidden for these agents, so a message sent while they worked always waited in the queue. The default choice shows the server's own default.
Gemini with a saved API key no longer stays on "Connecting" when Gemini does not answer its key sign-in. After the start time limit, the provider shows an error. Before, it waited with no end, and Cancel waited too.
The provider settings and the model picker show "Finish the sign-in if a browser page opens" and Cancel only while a sign-in waits for you. Before, they also showed when OpenBot checked a provider again after a key change, a restart or an update, when no browser page opens.
A right-click on the message box and on other text fields now shows Undo, Redo, Cut, Copy, Paste, Delete and Select All on Windows, Linux and macOS, and Paste and Match Style on macOS. Each item is available only when it can act, the labels use the app language, and a misspelled word shows its spelling suggestions. Before, the menu showed only Select All, or Copy and Select All with selected text.
Side threads now use your send shortcut setting.
A message sent while an agent works now goes straight to the queue. Before, it showed in the chat for a moment and then moved to the queue.
A message sent to an idle agent now stays in the chat. Before, it left the chat for a moment and the queue panel opened.
You can now sign in to the Make MCP server. Before, sign-in stopped with "The server answered 400."
The plan usage that windows, browser tabs and phones show is now read at most once every four minutes for each provider sign-in, however many of them are open. Before, each of them read it again, so a provider such as Claude could refuse more reads for a long time.
When a usage read fails, for example because the provider limits requests, OpenBot keeps the last reading and waits longer before it tries again: 5 minutes, then 10, 20 and 40, up to one hour. Before, the reading disappeared and OpenBot tried again every 5 minutes.
A server that you host yourself, such as on a VPS, now takes as many members as your license has users. Before, it stopped at 3.
Gemini no longer fills the memory of a computer or server over hours of use. Before, each Gemini chat and each model list refresh kept a helper process of about 100 MB until Gemini stopped, so a 4 GB server often held new messages for low memory. Now a model list refresh uses no more memory. On macOS and Linux, each Gemini chat also runs in a process of its own, and OpenBot ends that process when the chat has been idle for 10 minutes, when more than two chats are idle, or when memory is low. The next message continues the same chat.
An agent from the ACP registry, such as Mistral Vibe, now installs on Linux when its runtime archive puts ./ before each file name. Before, the install stopped with "The runtime archive contains an unsafe path."
A channel lead can now route a very long request. Before, a Grok lead got only part of a request of about 100,000 bytes or more, and the channel showed "The channel lead could not route this message". The lead now reads the first 20,000 bytes of the request.
Show all your agents when you go back to the mobile app. Before, agents that use DeepSeek, OpenRouter or Inception could go out of the list and lose their pin.
Server Settings > General > Timezone saves the zone that you choose. Before, only "Automatic" and "UTC" stayed selected, and a zone such as "Europe/Warsaw" did nothing.
The OpenBot Telegram bot now answers a mention such as "@meetOpenBot hello" in a group where it is not an admin. Before, only replies to the bot worked there, and a mention got no answer.
Settings now show Pi, Muse, Inception, DeepSeek and OpenRouter as off when you turned them off. Before, after a restart their switch showed on, and saving a key failed with "OpenRouter is off in OpenBot".
Put the cursor in the chat box when you click Reply. Before, after you went between an agent chat and a channel, Reply could leave the focus on the button, and you had to click the chat box before you typed.
The agent's browser can now click, hover, drag and upload files on a part of a page that it must scroll to. Before, these actions failed with "No node found at given location", for example when an agent attached a file to a GitHub comment.
Remote Desktop and device connections to a server connect through the relay when the server has more than one network, or many devices are connected. Before, the relay refused new connections after the first few.
A routine or provider handoff no longer cuts an emoji in half when it shortens long chat history. Before, the cut could make Codex reject the request.
Usage now shows the OpenCode Go quota when the OpenCode CLI has its own key: an OpenCode Console connection, a key from opencode auth login, or OPENCODE_API_KEY in the environment. Before, the quota showed only for a key saved in OpenBot.
Log in with code for Codex works when the stored ChatGPT sign-in can no longer refresh. Before, it failed with "Your access token could not be refreshed", or the code dialog closed after a second with "Connect ChatGPT to continue".
On a long page, a browser snapshot now shows the controls in view first. Before, controls earlier on the page could fill the snapshot, and the agent could not see the controls that you scrolled to.
When a member of a group chat fails because of a provider error, such as a rate limit or a spent quota, the group now shows the reason from the provider. Before, it showed only "The agent could not complete this task."
A transfer export keeps all the skills of a skill pack, such as pstack, as one skill. Before, it counted each skill of the pack, so an agent such as Dr. Eggbot lost skills at the limit of 32.
On a server that you joined, the secure card for a password or a code sends your answer to the agent again. Before, Submit, Take over and Cancel failed with http/400/team_api_error.
The mouse wheel and trackpad scroll in the remote desktop of a Mac. Before, the Mac did not scroll for most wheel steps and trackpad gestures. The Mac that you connect to must have this version.
A thread that you open in place of another thread now shows its latest message, not its first message.
Add account for Gemini on a joined server now opens the Google sign-in in the browser of your computer, as the first Gemini account does. Before, the page opened in the browser of the server, and Google refused it with "Your browser is not supported any more" (#4004). The server and the app must both have this version.
ChatGPT agents now run when Codex signs in with an OpenAI API key, also with codex login --with-api-key in a terminal. Before, OpenBot showed "Codex requires a ChatGPT subscription login" and did not start Codex. Connect now uses that login too, and does not open a ChatGPT sign-in.
DeepSeek models now work when an MCP server, such as Webflow, has a tool with a [ in a regex character class. Before, DeepSeek refused each request with a 400 error.
Agents find your 1Password and Bitwarden logins when you ask them to sign in with your password manager. Before, an agent did not find a tool for 1Password and looked only in the OpenBot Vault.
The $ list of a side question shows the agent's skills. It said "This agent has no skills."
The Local scripts guide now tells how to answer a question that lets the person pick several choices. The pending list shows these choices in the question text, and a script sends each picked label as one answer.
Codex agents now get the tools of a STDIO MCP server that has a working directory. Before, OpenBot left such a server out of Codex sessions, although its connection test passed.
Gemini agents in Full access can now read and change files in the folders that you add to them. Before, Gemini refused all files outside the agent's workspace and the shared folder, and Full access showed no way to add a folder.
A folder that you add to or remove from an agent now applies at the next message in an open chat. Before, Gemini and other ACP agents used the old folders until a new thread started.
Agents' browser typing no longer goes into a different field when the page moves focus. Before, the text or the Enter key could reach another field, and the agent was told that typing worked. Now the agent gets an error that says whether the text is already in the field.
A sign-in code email is sent again when the mail server does not answer. Before, OpenBot showed a "confirming delivery" countdown and then waited for a code that was never sent.
A channel task that you stop now shows Continue and Reassign. Before, it stayed stopped with no button. Tasks of an archived channel get the same buttons when you restore it.
"continue" in a channel now continues the work that you stopped last, also when older work is still open. The members get their original request again, not only the word "continue".
A channel message that starts with a typed "@Name" now goes to that member directly.
When the lead cannot read its own routing answer, the channel now says so and offers Reassign.
The lead no longer posts "Assigned to" itself when it answers a message alone.
In the remote desktop, Cmd+A or Ctrl+A and the zoom keys (Cmd or Ctrl with =, - and 0) now go to the remote computer when the viewer has focus. Before, OpenBot kept these keys: select all did nothing and the zoom keys changed the OpenBot interface size.
A Cursor agent can use OpenBot's tools and message its teammates when openbot is in Cursor's list of disabled MCP servers for the project. Before, Cursor started the agent without OpenBot's tools and showed no error, so messages to teammates were not sent.
When Cursor does not load OpenBot's tools for another reason, such as a Cursor team policy, the chat shows an error. Before, the turn ended and the model reported tool errors that did not occur.
Keep text from later browser frames in snapshots after the control limit is reached.
Agents no longer load a tool schema again when the tool search already returned it.
A channel no longer shows old replies again as new messages after OpenBot starts again.
A channel turn that you stop no longer leaves an empty reply.
When a request goes to several members, each member now gives its own view. Before, a member could copy the reply of a member that answered first.
Members no longer get notes such as "Result submitted to the shared chat." from other members, and a member now ends its turn when it posts its result.
Channel tasks that you stopped in an earlier version now show Continue and Reassign after the update.
Agents now call a tool directly when tool search already returns its schema. Before, the agent instructions told them to describe each tool before its first call, which was in conflict with the search results.
Agents can now open a page that goes to its next page while it loads, as sign-in pages often do. Before, the agent got an ERR_ABORTED error, and the open tool closed the new tab.
Cursor and other ACP agents no longer ask you to sign in again after a timeout, a network failure or a server error during start while you are still signed in. Only a real sign-out, such as an "Authentication required" answer or 401 Unauthorized, asks for sign-in. The log now keeps the text of the agent's answer.
The agent's avatar in the sidebar now shows that the agent works when it works in a side thread. Before, the avatar stayed still until the agent worked in its own chat.
When a browser action timed out after its click or key press reached the page, the agent got only "timed out" and could repeat the action. The error now says that the input reached the page and tells the agent to check the page first.
You can now drag the window by the empty top row of the settings page.
A custom endpoint on a llama-swap server no longer unloads the agent's model at the start of a new chat. OpenCode asked its small_model for a chat title, and llama-swap loaded that model, then loaded the agent's model again. OpenBot does not use these titles, so it now turns them off.
Computers, the web client and phones now use Team API protocol 9, which has room for more features. A server on an older version and an app on an older version keep working with this version: they do not get the features that need protocol 9, such as notification levels that apply on all devices.
The iPhone Live Activity that the computer updates now hides secrets and email addresses, as notifications do. It also no longer offers Approve for a command whose secret it hides, so you approve such a command in the app.
Your phone now gets the notification of an agent event that came just after you left the OpenBot app. Before, the computer did not send it when the app went to the background before it told the computer, and the notification was lost.
An iPhone Live Activity now takes the place of a notification only when it shows that event. Before, a reply or a failure did not notify while the activity showed another agent at work or a request of another agent.
An iPhone Live Activity that does not get its updates no longer stops notifications. The computer also sends a notification again when Apple or Google do not take it the first time.
An iPhone Live Activity now takes the first update from the computer after you leave the app, and ends also when the phone was without network at that time. Before, it could stay out of date.
A desktop notification for an agent on another server now shows also when the list of its agents cannot be read at that moment.
Picture in Picture now shows the page of a joined server's browser. Before, the floating window stayed empty and dark.
Picture in Picture now follows the tab that the agent changes to on a joined server.
Keep the web app on screen when openbot.run gets an update. Before, the page could go blank and load again after an update.
Connect the web app to your server again within about 30 seconds after the server comes back. Before, the web app could wait up to 2 minutes, and you had to reconnect from Settings.
The web app now connects again at once when you go back to its tab or window. Before, after some failed attempts it could wait up to two minutes, also when the server was already back, and you had to reconnect from Settings.

Oct 29, 2026

Yaji 0.54.0

New

A Workspace only agent can now change files in folders that you add. Open the agent's settings, then Folders, or drop a folder on the chat. The agent knows which folders it can change.
You can set the context window of each model of a custom provider on this computer. OpenCode then compacts the conversation at the correct size, and does not use its 200K-token default.
Ctrl+F (⌘F on macOS) now searches the open side thread. Before, the shortcut did nothing in a side thread.
The website has a Pricing page. It shows the hosted servers at their Stripe prices, the Starter license (1 member) and the Team license (2 to 10 members), monthly or yearly, in EUR, USD or PLN. Choose a server, a license or both. The summary shows the total. Visitors in the EU see that the prices do not include VAT. A license bought with a server costs 10% less (bundle discount), and yearly billing takes 20% off a license. When the Stripe prices do not load, the page shows the list prices.
The header and the footer of the website have a Pricing link.
The Pricing page has one button. It opens the web app at the add server dialog, or at Billing when you choose only a license.

Improved

An agent's post to Slack, Discord or Telegram now gives the agent the ID of the post, and for Discord a link to it, so the agent can link to the post.
A new server has fewer members: Starter 1, Standard 5 and Pro 10. Servers that already exist keep their limits (Starter 3, Standard 10, Pro 25).
Sell the commercial license in two plans. Starter is €10 a month for 1 user. Team is €25 a month for 2 users and €15 for each extra user, up to 10. For more than 10 users, contact us. Yearly billing takes 20% off. An account that pays for a hosted server (any plan) then pays 10% less for the license. Each user can use up to 3 devices.
On the Pricing page, set more than 10 members for the Team license to see the custom price and a Contact us button.
Add server lets an account with no license choose personal or work use. For work, set the people at work, and buy the server and its license at the bundle price with one payment. The license cannot have fewer users than the people who use your servers now, and the order uses the currency of your other plans.
Buy a license alone from Billing: set the people with a slider and see the price at once.
Change a Starter license to Team, or a monthly Team license to yearly, from Billing. When an invite needs more than 1 user, the members section offers the change to Team.
Make a new commercial license free until 1 December 2026. Billing, Add server and the pricing page show "Free until 1 December 2026". Stripe takes the first payment on that day. Cancel before then to pay nothing.
The pricing page sends its order to the web app. After sign-in, the app goes straight to Stripe Checkout with the same server, license, users, billing period and currency, so you do not choose again. When the order does not fit the account, such as another currency than your other plans, the app shows the order filled in and you confirm it.
Update the Privacy Policy, Terms & Conditions and Commercial License Agreement on openbot.run. Add the Data Processing Addendum at openbot.run/dpa. openbot.run/privacy shows the Privacy Policy too. openbot.run/commerciallicense and openbot.run/branding open the existing pages.
The site footer links to the Terms, the Commercial License Agreement and the DPA.
When you add a server for work, the dialog now tells you why it does not let you choose fewer people: everyone who uses your servers now needs a seat on the license.
Prices in PLN are 4 times the EUR price, so they are round amounts. The Starter license costs 40 zł a month, and the Team license costs 100 zł a month for 2 users and 60 zł for each extra user. A new hosted server costs 80 zł (Starter), 200 zł (Standard) or 400 zł (Pro) a month, not 90, 220 or 440 zł.

Fixed

OpenBot starts on a Mac where /Library/Application Support or /Library/Application Support/OpenBot has a non-root owner or group write permission, and no host manager is installed. Before, OpenBot stopped at each start with an empty error. A start error from the host manager check now shows its cause.
Remote desktop pairing on macOS 27 works again. Before, the host stopped at each pairing attempt and every setup check showed as not checked.
When an agent cannot open a secure password or code card for a sign-in, it now gets the reason, such as which field or button target is wrong, and can correct the request. Before, it got only "Secure authentication is unavailable".
Routine agents now check saved Vault logins before they report a sign-in requirement. They can use saved passwords and authenticator codes through secure browser input.
Quit OpenBot without the error "OpenBot database is not initialized." when an agent reports its context use during the quit.
Keep a self-hosted or hosted server responsive after an error in the main process, and let it finish a quit. Before, an error box that nobody could close on the server display stopped the process, and the service manager did not start it again. On desktop, an error during a quit also shows no error box.
OpenBot keeps working when a page in the browser requests an address that it cannot read, such as a host that starts with an escaped space. Before, OpenBot stopped and did not answer until a restart.
Show the reason of a group chat task that stopped because its member was removed or is not available in your language. Before, the reason was always in English.
Show only Reassign on a stopped group chat task whose member left the group chat. Before, Continue stopped the task again.
Provider settings in agent settings show in one card, as the other groups do. Before, each switch showed under its name and each choice was a full-width field outside a card.
Keep a new provider account, such as another Gemini account, when its sign-in works but its first usage reading fails. Before, OpenBot removed the account after the Google sign-in, so it looked as if OpenBot did not add the account. A Reconnect with the same failure now shows the account as connected, not as an error.
Keep the chat steady when you switch between agents that have queued messages. Before, the queue animated the other agent's messages out and this agent's messages in, and the chat jumped while the queue changed height. When the agent that you open has no queued messages, the queue of the other agent now closes at once.
Let Gemini continue when one MCP server that OpenBot gives it does not start, such as a server that asks for a sign-in. Before, each Gemini turn failed with "The MCP server '<name>' failed to initialize". Now OpenBot leaves the server out, tells you which server it left out, and sends the message again.
Cancel and × now stop a browser sign-in in the Apps connect dialog. Before, both stayed off while the dialog waited for the browser, for up to 5 minutes.
Show the agent's steps in a side thread, as in the agent's own chat: the thinking while the agent works, and the steps row above each answer. Before, a side thread showed only the answer.
On macOS, OpenBot → Check for Updates… opens Settings on the Updates tab, where the check and its result show. Before, nothing changed on screen when OpenBot was up to date or the check failed.
Check the remote desktop permissions when you open Server settings → Remote desktop on a Mac. Hide the Grant button of a permission that is already allowed. Before, all rows showed "Not checked" and Grant until you selected Check again.
Cursor turns fail less often with "Invalid model value". Cursor asks its server for the model list each time the model is set, and reports a failed request as an invalid model. OpenBot now sets the Cursor model only when it changes, not before every turn (#3800).
An agent that a ChatGPT (Codex) usage limit moved to a backup model now goes back to its model within 15 minutes when Codex ends the limit early, for example after you spend a reset credit in the Codex app. Before, OpenBot kept the agent on the backup model until the old reset time, also when you selected the model again.

Oct 28, 2026

Yaji 0.53.2

Action needed after you upgrade

On Windows, OpenBot updates again when it was installed in %USERPROFILE%\OpenBot, the folder that also holds your agent folders. Before, each update stopped with "Failed to uninstall old application files. Please try running the installer again.: 2", also after a restart. The update now installs OpenBot in %LOCALAPPDATA%\Programs\OpenBot and removes only the app files from the earlier folder. Your agents, Shared and Downloads stay where they are. A taskbar pin of the earlier copy stops working: pin OpenBot again from the Start menu. The Start menu and desktop shortcuts, the openbot:// links and the open-at-login setting move to the new copy. Do not uninstall OpenBot from Windows Settings to repair this: in that folder, the uninstaller of 0.53.1 or earlier also deletes your agent folders.

New

Agents now save lasting preferences and facts from your chats and side threads in the background. After every 20 turns, after 15 minutes with no new turn, when you clear a chat and when you close a side thread, the agent's provider reviews the new part of the conversation and adds, updates or forgets memories. The conversation shows the changes. Memories that you wrote or set stay as they are.
Agents now search their saved memories before they say that they do not know something about you. A search also finds other forms of a word, so "closing" finds "closes".
A memory review uses the provider's smaller default model when the agent's model costs more, such as Claude Haiku for an agent on Opus or Sonnet.
Server settings > General > Memory review sets how often the agents on a server review their memories: every 10, 20 or 40 turns, or Off. Off also stops the reviews after a pause, a clear and a closed side thread. An owner or admin can set it for this computer and for a joined, hosted or self-hosted server that runs this release.

Improved

An agent can now keep 256 memories by default instead of 64, so the background reviews have room for many small facts. Only the essential memories go into each prompt, so this adds no cost. A limit that you set in Server settings > General > Memories per agent stays as it is.

Fixed

The Windows installer no longer accepts %USERPROFILE%\OpenBot, %APPDATA%\OpenBot, or a folder that holds one of them, as the install folder. It uses the default folder instead.

Oct 28, 2026

Yaji 0.53.1

Improved

Server settings › Storage no longer shows the disk, memory and CPU use of the machine or its alert levels. OpenBot no longer sends machine alerts and does not show the machine card in the sidebar.

Fixed

While a question, approval or browser request waits for you, the activity line under the chat now says that the agent waits for your answer, your approval or you in the browser. Before, it kept the last tool step, such as "Working through the next tool-assisted step…".
A message that you send with a Markdown table is now as wide as an agent table. Before, the message stayed 420px wide, and the table broke names and dates in the middle. In a narrow window, the table scrolls sideways.
The model picker shows the Gemini models also when Gemini starts slowly. Before, Gemini had only 5 seconds to list its models, and the picker showed "No models are available from Gemini" until the next refresh.
Queued messages no longer show Steer for Gemini, OpenCode, Cursor, Cline and custom agents. These providers cannot take a message into a running turn, so Steer always failed. The message still starts when the turn ends.
On macOS, the blue Computer Use rim no longer shows on top of the OpenBot window when that window is in front of the agent's window. Before, OpenBot read the windows in the wrong front-to-back order on some macOS versions, and did not cut the rim where the OpenBot window covers it.
A failed update check, download or install now writes its reason, error code and HTTP status in logs/main/main.log. Before, the logs showed only check_failed, with no reason.
The activity log in openbot.db now records the turn of each turn.completed event and keeps the turn of each activity when OpenBot rebuilds a thread. Before, these rows had no turn.
In a workflow Agent or LLM step, a long model name now stays inside the field and ends with "…". Before, the model button went past the right edge of the field. A long agent name next to it stays full until the model button has only room for a few letters.
Agent chats in the sidebar now show their new replies when the app starts before your account finishes signing in. Before, the unread counts of those chats were missing after the start.
A group chat no longer shows its lead as "Working on it…" while the lead is paused. Before, a message that waited for a paused lead showed the lead at work until the user restored it, but the lead did no work.
Usage now shows an estimated cost for Gemini (Antigravity) turns that use Gemini 4 Argon, Gemini 3.8 Flash, Claude Sonnet 5.5 or Claude Opus 5.5. Before, these turns had token counts but no cost. The estimate uses the new Gemini rates from the day that Google changes them. Turns from before the upgrade stay without a cost.

Oct 28, 2026

Yaji 0.53.0

New

Server settings › Storage shows the disk, memory and CPU use of the computer that runs the server, for owners and admins. It updates each minute while it is open.
Set an alert level for disk, memory and CPU, and choose the owner or admin who gets the alert. When the use passes a level, that person gets a notification on each of their computers, and again each day while it stays above the level. A server must have this version to use it.
While a level is passed, owners and admins see a card at the bottom of the sidebar with the use and who gets the alert. It opens Server settings › Storage.
A side thread shows how full its context is next to the send button, as the agent's own chat does. Open it to see the details, and select Compact to compact the thread's context now. Compact shows for providers that OpenBot can ask to compact, on this computer only.
A thread that you read on this computer is read on your phone too. The server keeps your read place in each thread for all your devices.
You can send files in a thread of a joined server and in the web app. The server must have this version of OpenBot.
Add guides for importing Grok Bot agents and Hermes Agent profiles into OpenBot, with screenshots and instructions for each import method.

Improved

The Timezone list in Server Settings has a search field. Type part of a zone, such as "new_york", or its name, such as "Eastern", then use the arrow keys and Enter to choose it.
When an agent answer names an output file by its full path and does not attach it, OpenBot now attaches the file to the answer. This applies to images, audio, video, ZIP, PDF, office documents, CSV, text and Markdown files of 25 MB or smaller in the agent's workspace or the Shared folder. Before, the iPhone app and other devices showed only the path, and you could not open the file. Code, configuration, log and SVG files, and folders, stay as paths.
The "Add backup model" list in agent settings has a search field. Type part of a model name, a provider name or a model id, such as "glm" or "openrouter/qwen", then use the arrow keys and Enter to add the model.
On a sidebar of 300 px or less, the wait chip of a "Needs you" row shows only its icon, so that a long agent name has more space.
Agent avatars and photo avatars have a thin outline in the colour of the row behind them. In a group avatar, the faces are larger and overlap, a photo has the size of the other faces, and each outline cuts the face below it. The count of more members is a round disc that overlaps the last face. A count of two digits goes under two faces.
The OpenBot server offer opens as a window only one time: after your first routine on this computer. Before, it opened after a new routine on this computer, a remote computer or the web app, up to three times. The note above the routine list and the line in the offline notice do not change.
Settings › Mobile Connect shows your phones as cards. "Add phone" opens a short dialog that asks if OpenBot is on the phone, then shows only one QR code: the App Store or Google Play code, or the sign-in code. A tag under each code says what it does. Before, the tab showed the download code and the sign-in code together, and a phone camera could scan the wrong one.
The permanent link is a card among your phones. Its code opens in a dialog with a warning, and hides after one minute. Replace and revoke are in its menu.
"How devices connect" is a button in the tab header that opens both choices.

Fixed

Sliders, such as the Dynamic Island width and height in Settings, move again with the pointer and the arrow keys. Before, they did not move.
OpenBot starts after an upgrade when its database has broken links between rows from an earlier version. Before, the upgrade stopped at each start with "OpenBot database migration to version 47 failed", and the app did not open. OpenBot keeps these rows as they are, and an upgrade step stops only when it breaks a link itself.
A custom provider on an Azure OpenAI or Azure AI Foundry v1 endpoint (https://<resource>.openai.azure.com/openai/v1) runs agent turns in OpenCode. Before, each turn failed with "Function tools with reasoning_effort are not supported".
The $ list in an agent thread shows the agent's skills and MCP servers. Before, it said "This agent has no skills.", and a skill in a sent thread message showed as unavailable.
You can paste, drop or pick images and files in a side thread of an agent on this computer, and they go with the thread message. Before, a pasted image in a thread made a sound but went to the draft of the agent's main chat. A thread of an agent on another computer takes text only, and now says so.
A chat in "Needs you" shows the question or the command as its preview. Before, it showed the previous message, and the question was only in the tooltip.
On a server that you joined, a tab that you open in an agent's browser panel opens in the agent's browser profile. Before, the tab opened in the shared profile, so a sign-in there was available to other agents.
When an agent asks a question in a thread, the thread now shows the question card and waits for your answer. Before, the question showed only as text, and you could not answer it.
While a side thread waits for your answer to a question or an API key request, the thread and its tab now show that it waits for you. Before, they showed that the agent worked.
When a side thread's agent asks you to take over the browser, the thread and its tab now show that the thread waits for you. Before, they showed that the agent worked.
When your answer or approval in a side thread fails, the error now shows where you answered. Before, it showed in the agent's own chat.
In the web app, and on a computer joined to a server, a side thread now shows the API key card and the other question forms. Before, it showed only a plain question. The server must also run this version.
On Windows and macOS, OpenBot now opens the Google sign-in page for Gemini, as it does on Linux. Before, the Gemini server opened it itself, and when no page opened, Connect showed no error and waited ten minutes.
When the Gemini process stops answering, OpenBot now ends it and starts a new one. Before, a turn could wait with no end, and Gemini worked again only after you restarted it.
The channel header shows the face of each member. Before, the faces were almost on top of each other, so the header showed only one face.
When an MCP server that you added in Server settings needs a sign-in, the agent now shows the Sign in card in the conversation. Before, the card showed only for a server that the agent added.
In Discord, the agent now gets the text and files of a forwarded message, when you reply to the forward or mention OpenBot after it. Before, the agent got only your mention and said that it saw no report. When Discord does not give the forwarded content, the agent is told so.
The activity line of an agent that uses the built-in browser now says what the step does, such as opening, reading, clicking or typing on a web page. Before, each browser step showed "Searching for current information…" and "Reviewing the sources and information I found…".
About 30 seconds after iOS puts the OpenBot app to sleep, a server now updates the iPhone Live Activity and Dynamic Island, and sends notifications. Before, when the app went to sleep before it told the server, the activity could show "Working" for more than two minutes after the agent finished, and notifications for that time did not come.
An agent on a second provider account uses that account when it routes group work, makes a group summary, answers a side question, or writes its profile. Before, these requests used the provider's first account. When the first account was at its usage limit, a group lead on the second account stopped with a usage limit notice, and the usage limit card did not offer its model.
A chat with unread messages now keeps the "NEW" divider in view while its images and previews load. Before, content that loaded after the chat opened moved the divider down, and the chat showed older messages.
On Windows, Remote Desktop starts again. The remote desktop runtime did not include the Sunshine DirectX shaders, so Sunshine did not start and OpenBot showed "Sunshine did not start on the host".
On macOS, the name of the image in the image viewer no longer shows under the window controls. Before, the first letters of the name were behind the close, minimize and zoom buttons.
On macOS, OpenBot can now ask for Automation access, so the Apple apps and Messages tools work in the released app. Before, macOS refused the request without a prompt, and a Messages lookup stopped with bootstrapTimedOut.
When Gemini starts slowly after its first download, as on an Intel Mac, OpenBot now keeps the download and tries again, and waits twice as long at each try. Before, OpenBot deleted Gemini after 3 minutes, and each Retry downloaded it again and failed the same way.
After you clicked a channel in the sidebar and then went back to the OpenBot window, the avatars on that row moved without stop, also when the agents were idle. Now they move only while the pointer or the keyboard focus is on the row.
On a computer joined to a server, the chat of the server's agent now shows the context window meter beside the microphone. Before, the meter was missing. The server must also run this version.
Start OpenBot when a saved MCP server has headers or other values that it cannot read. Before, the app did not open until the server was deleted from the database. Now Server settings shows that server turned off, and you can correct it or remove it there.
When an agent started a new ChatGPT session, for example after it moved to a backup model, the chat could show a red paginated_threads is not supported yet error. ChatGPT cannot read a session before it saves the first turn, and OpenBot now does not show this as an error.
Show the Download button for an image attached in a side thread, under the image and in the image viewer. Before, only the agent's main chat had it.
Show the Download button for an image file that an agent message names by its path, under the message and in the image viewer. Before, only an attached image had it.
Usage now shows an estimated cost for Codex turns that use GPT-6.1 Terra. Before, these turns had token counts but no cost. Turns from before the upgrade stay without a cost.
An MCP server that refuses to register OpenBot, such as Dropbox, now shows Refused and says that the server does not accept OpenBot as an app. Before, it showed Failed with the server's own text.
An agent can attach files in the built-in browser when the page opens its file chooser from a button, such as the "Paste, drop, or click to add files" button in a GitHub comment. Before, upload_files worked only on a file input, and on these pages it stopped with "Node is not a file input element".
A channel no longer stops all work when a member repeats its result as its last reply. Before, that reply could stay "still arriving" after the turn ended, and every later assignment in the channel failed with "The provider could not start this assignment". A channel that this blocked works again when OpenBot starts after the upgrade.
A second Gemini account stays after its Google sign-in when Gemini starts slowly, as it can on Windows. Before, OpenBot stopped the check after the sign-in at 30 seconds, and the new account disappeared.
When a new provider account cannot sign in, OpenBot now shows an error with the reason. Before, the account disappeared with no message.
Cancel on a provider account that is still connecting no longer waits until the provider finishes starting.

Oct 27, 2026

Yaji 0.52.1

New

Remove the downloaded CLI of a provider that is off. Open the provider in Settings > Providers and select Remove download. It shows the space it frees. Accounts, sign-ins and agents stay. When you turn the provider on again, OpenBot downloads the CLI again (#3321).
An owner or admin can import a CSV password export into a server's Vault from the iPhone app. The server answers only how many logins it added and skipped.
An owner or admin can set the default provider of a server from the phone. The server starts new agents on that provider. Clients older than Team API protocol 8 do not see DeepSeek, OpenRouter or Inception as the default.
The Apple apps listing in the Marketplace has a switch that asks macOS for access to Calendar, Contacts, Mail, Notes, Reminders, Messages and Safari when you install it. macOS shows one prompt for each app, so the prompts no longer stop an agent in the middle of a task. After the install, "Ask for access" asks again, and a message names each app that macOS did not allow. macOS only.
The Vault keeps API keys. Add a key with its variable name, such as OPENAI_API_KEY, in Server settings › Vault › New item › API key. Every agent on the server finds it with list_api_keys and reads it from a private file in its commands, so the key does not go into the chat. Only the user can add, change or delete a Vault API key. An owner or admin of a joined server can manage its API keys from the desktop app or the web app when the server has this version.

Improved

Agents get shorter instructions, with each rule in one place, so each turn uses less context.

Fixed

Show Mermaid diagrams that have a line break in a label. Before, the chat showed a broken image.
The secure password card now always shows "Save to Vault" and names the Vault: this computer or the joined server. An owner or admin of a joined server can save the password to that server's Vault; the box starts unchecked. A member, or a server that needs an update, sees why the box is off. Before, the option did not show on a joined server or in the web app, and the password was not saved.
While an agent writes, the activity line shows its text without Markdown marks such as ** and - .
Start Claude agents as new sessions when a Claude Code session opens OpenBot, for example with open -a OpenBot. Before, the agents got the session ID, entry point and messaging token of that session, and each new chat wrote its prompt cache again.
Let Gemini start MCP servers whose command, such as node, is on the login shell PATH. Before, OpenBot opened from the Dock or a login item could not find the command.
OpenBot now quits when a file operation never finishes, for example when macOS waits for a privacy approval that nobody answers. It stops itself at most 10 seconds after its shutdown, which can take up to 30 seconds. Before, OpenBot stopped responding after you quit it, and it did not open again until you forced it to quit on macOS or ended it on Linux.
Agents now follow a confirmation rule in any CLAUDE.md or AGENTS.md file, also when a request names the action. Before, an agent could make a change that cannot be undone, such as a Stripe setting, without the confirmation that the file asked for.
When an agent asks for a password, a code, or browser control in a thread, the thread now shows the card. Before, the card did not show, and the agent waited without end.
An OpenCode agent's answer after a background subagent ends now shows in the chat. Before, OpenCode saved the answer, but the chat never showed it, because no message had started that turn.
An OpenCode answer that the agent writes in two steps now shows as two paragraphs. Before, the text of the two steps showed as one line.
Text that you type in a side thread and do not send now stays in the message box of that thread. It stays when a turn in another thread of the agent ends, when you open another thread or chat and come back, and after a restart. Before, the text was removed when a turn of the agent ended.
Connect Railway works again. OpenBot now registers as a public client when the sign-in server supports it. Before, Railway refused the client secret that it gave, and the sign-in stopped with "Existing OAuth client information is required when exchanging an authorization code".
Show a clear message when a sign-in server refuses OpenBot at the end of an MCP sign-in. Before, the message did not say what went wrong.
When this computer's server was published, OpenBot publishes it again after a start, also when the account service or Signal does not answer at first, for example at login before the network is ready. It also publishes the server again when Signal disconnects it later. It tries again after 1 minute, and then less frequently. Before, the server stayed private after one failed try, and you had to publish it again, for example after an update.
A long context compaction with Codex or Pi now runs to its end. Before, OpenBot stopped each compaction after 2 minutes, so the compaction of a long thread did not complete. OpenBot now stops a compaction only when the provider does not start it in 2 minutes, or when it runs for 20 minutes.
When Gemini or another ACP agent crashes, the error now shows the first line of the crash report, such as F… Check failed: … or *** SIGABRT received …. Before, it showed only the last line, RAW: Raising signal 6, which does not say why the agent stopped.
End a Remote Control session when its stream fails and you go back to OpenBot. Before, the session stayed open, and the Remote Control button showed a green dot on every agent of that server.

Oct 27, 2026

Yaji 0.52.0

New

Add Obsidian to the Apps catalog. Agents can search, read and write the notes in an Obsidian vault, so the vault can be their memory. Install the Local REST API plugin in Obsidian, turn on its HTTP server, and type its API key in the connect dialog.
Add Neon, Prisma, Netlify, Railway, Atlassian (Jira and Confluence), Todoist and Hugging Face to the Apps catalog. Sign in in the browser to connect each of them, except Hugging Face: it reads public Hub data and needs no account. Neon, Prisma, Netlify, Railway, Atlassian and Todoist give agents write access, so agents can change or delete data there. Each listing says what agents can change.
The website has the OpenBot Commercial License Agreement at openbot.run/commercial-license. The Terms & Conditions link to it.
Give an agent more than one conversation. Click + in the chat header to start a new thread. Each thread has its own tab and runs at the same time as the other threads of the agent. The agent keeps its main chat.
Rename, close, archive and delete threads from the tab menu. An archived thread moves to "Archived" in the thread list, and opening it restores it. Delete shows an Undo button for some seconds first.
Use threads in the web app and on a computer that is joined to a server, when the server has this version.
Search on this computer also finds messages in open threads. Select a result to open its thread tab.
A notification from a thread opens that thread's tab.
An image file that an agent names in its answer, such as a PNG or SVG in the shared folder, now shows as a preview under the answer. Click it to open it at full size.
Right-click an agent in the sidebar and choose "New thread" to open its chat with a new thread.
Attach MP4 videos in chat. The chat plays them, and the agent gets the original file. A remote host must run this version of OpenBot to accept MP4 files.
Add Render to the Apps catalog. Connect it with a Render API key. Agents can list workspaces and services, read logs, events and metrics, find why a deploy failed, trigger deploys, and run read-only SQL on Render Postgres. The key reaches every workspace of the account, so agents can also create services and databases and replace the environment variables of a service. The server cannot delete a resource.
An owner or admin of a joined server can now open that server's Vault in Server settings, in the web app and in the desktop app. They can add, change and delete logins, cards and contact info, and show or copy a saved password, card number or security code. A CSV import is available only on the server's own computer. Before, the Vault showed only for this computer, so the web app had no Vault.
Add Trello to the Apps catalog. Sign in to Trello in the browser and choose the workspaces. Agents can search and read boards, lists, cards, checklists and comments, create boards and cards, move lists and cards, and mark cards done. They can also read and add cards in your Trello Inbox and read your Planner events. Agents have the same permissions as you in the workspaces that you choose. They can archive cards and lists, but the server cannot delete anything permanently.
Dutch interface. Choose Nederlands in Settings > General > Language, or keep System default on a computer that uses Dutch. The web client uses it too.

Improved

Update the Terms & Conditions on openbot.run/terms. They now describe free non-commercial use, Commercial Licenses (Starter and Team seats), hosted server plans, the refund and withdrawal rules, 30-day hosted server data retention after expiry, and Marketplace content reports. The minimum age is now 16.
Change the minimum age in the Privacy Policy to 16, the same as the Terms & Conditions.
In the web client, the Marketplace now opens as a full page of the window, as in the desktop app. Before, it opened as a dialog over the workspace.
The Marketplace now has a side panel, as the settings page has. Select Apps, Agents or Skills there. Below them, "Your apps" lists the apps that you connected, and a warning shows on an app that needs your attention. While you search, the panel shows how many results each kind has.
The Marketplace now opens on Apps. Apps show in groups by category, in compact rows with Connect, Connected or Reconnect at the end of the row.
To add your own MCP server, click "Add custom" next to the search on the Apps page of the Marketplace. Server settings open on the form for a new MCP server. Only a user who can add MCP servers to the server sees the button.
On an app, agent or skill page of the Marketplace, the top row now shows the path to it, for example "Apps › GitHub". Click "Apps" to go back to the list. On an app, agent or skill page, Apps, Agents or Skills stays selected in the side panel. On the page of an app in "Your apps", that app is selected. The page has the same width as the list, so the content does not move when it opens.
Connectors are now only in the Marketplace. Server settings no longer has a Connectors section. To connect or change Slack, Discord, Telegram, GitHub, 1Password or Bitwarden, open Marketplace › Apps.
All app pages now use one layout: the header, then alerts, then sections with a heading and one card each, and last the danger zone. Text, forms and notes now show inside the cards, and all buttons in a card have the same small size. A plugin that is not completely installed shows a warning above its sections. An app that is not connected now shows "Not connected" at the top of its page. Before, some pages showed "Not set up".
Slack, Discord and Telegram pages now have a danger zone that disconnects every workspace, server or chat at once, after you confirm. Each row still has its own Disconnect.
The page that opens after a Slack or Discord sign-in in the browser now tells you to open Marketplace › Apps to try again.
The connect dialogs of Marketplace apps that you sign in to, give a key to, or turn on in their desktop app now look and work as the GitHub and Slack connect dialogs do. They show OpenBot and the app with a line that moves while they connect, and Cancel and the main button at the bottom right. When a connection fails, the line breaks. Connect in the Marketplace now starts the connection at once: a sign-in opens the browser, and an app on this computer connects if its server is on. An app that needs a key opens with the cursor in the key field.
To disconnect Bitwarden, use the button in the new danger zone at the bottom of its page.
The Cloudflare, Airtable and Paper icons in the Marketplace are now centered and clear on a dark background.
A plugin page now has a Connection section. It shows how each app connects: you are signed in, an API key is saved, it uses a desktop app on this computer, or it needs nothing. Use "Sign in again", "Change key" or "Connect again" there.
When the sign-in of a plugin expires, the plugin shows "Needs attention" and a warning, and "Sign in again" signs in again. The page header, the Connection row and the Apps list use the same name for this action. A plugin with more than one app shows the button only on the row of the app that needs it.
Every connector page now has an Information section with the developer, the category and the website.
On the Slack, Discord and Telegram pages, the sections are now named "Workspaces", "Discord servers" and "Chats". "Add workspace", "Add Discord server" and "Link a chat" are at the right of that section.
The 1Password and Bitwarden pages now look the same. Bitwarden shows the shared folder and the server address in one card, and the session key field has a label.
The danger zone of every app page is now named for the app, for example "Disconnect Notion".
Under the name of Slack, Discord, Telegram, GitHub, 1Password and Bitwarden, the page now shows a short line, as the Apps list does. The full description is at the top of the Information section.
On the Bitwarden page, the shared folder row now says "Folder in Bitwarden" under the folder name.
The developer of each catalog app now shows as a name, for example "Notion", not as a web address.
Tasks are now threads. A task that an agent starts opens as a tab in the chat header, and your existing tasks show as threads after the upgrade. Closed tasks show as archived threads. The Tasks page and the task panel are removed.
"Tasks at once" in the agent settings is now "Parallel threads". "Run as task" on a queued message is now "Run in new thread".
Deleting a thread also deletes its messages and their files from the agent's inbox.
After this upgrade, an earlier version of OpenBot cannot open your data.
A message from another agent now steers the running turn when the agent that gets it is set to steer busy messages. Before, it always waited in the queue until the turn ended.
You can steer a teammate's answer into the running turn with Steer in the "Waiting for replies" block. Before, the answer waited for the turn to end.
The iPhone app is on the App Store. Settings > Mobile, the download pages and the changelog page link to the App Store, not to the TestFlight beta.
Messages that you send while an agent works now start one next turn together, in their order. Before, each message started its own turn.
Slack, Discord and Telegram conversations send the agent's messaging guidance and the saved facts of the server only in the first turn of a provider session, and again when a fact changes. This makes each later turn smaller.
The agent's name in the chat header opens the agent's main chat while a thread is open. In the main chat it opens the agent settings, as before. The main chat no longer has a tab, and the thread list shows only threads.
Agents get a shorter list of models when they look for a model for a teammate: the default model, the models that agents use, and the first models of each provider. They can search for the other models. This uses less of the agent's context with providers such as OpenRouter that offer hundreds of models.
In the desktop app, when Cursor uses a merged plan, such as an X subscription, the sidebar shows one usage ring for Grok and Cursor. Before, both rings showed the same shared quota. Hover the Grok ring to see both.
A server now offers DeepSeek, OpenRouter and Inception, as the computer does. An owner or admin saves the provider's API key in the server's provider settings, and the server downloads Pi when it needs it. Phones and team members also see agents on these providers. Clients older than Team API protocol 8 still do not see them.

Fixed

A Cursor agent with Workspace only access on macOS can sign in again. Before, each of its turns stopped with "Cursor did not return an authenticated account." although Cursor was signed in.
Settings › Profile now says that its activity is the activity on this computer. Before, the title showed only "Activity", also when a different server was open.
When you save, change or remove a custom model endpoint while an OpenCode task runs, OpenCode now restarts after its tasks stop, and the new models show in the model picker. Before, OpenCode kept the old models until it stopped for another reason.
On Windows and Linux, the Alt key no longer shows the menu bar. Before, a single Alt press, or a keyboard layout switch with Alt+Shift, moved the page down and took focus from the message input. Menu shortcuts work as before.
Reconnect on a plugin that is already connected no longer fails with "An MCP server named … already exists". It now updates the saved connection.
A Cursor reply no longer ends with "Error: RetriableError: WritableIterable is closed" after a turn that succeeded.
When an agent asks another agent from a side thread, the answer now comes back to that side thread. Before, it arrived in the agent's own chat. The request still arrives in the other agent's own chat.
A side thread now names the other agent as the author of its answer. Before, screen readers read the answer as a message from the thread's own agent.
A chat with queued messages no longer slows down OpenBot. Before, each queue change wrote the queued messages to the database again and sent the whole chat to each connected device, and the queued messages went out of the chat and came back.
Text that you select in a code block of a message shows the selection actions (Quote, Explain, Improve). Before, a double-click or a selection in a code block showed no actions.
OpenBot starts again when an update for a provider CLI is available. Before, the app stopped at start and showed no update notification.
Muse agents on Windows can start a turn again. Before, each turn failed at once with no reply, because OpenBot sent the workspace folders in a form that Muse does not accept.
A Pi agent keeps its OpenBot tools after OpenBot restarts. Before, a chat that was opened after a restart lost every OpenBot tool, and each restart gave the tools new names, so the agent could call old names and get "Tool ... not found". A Pi chat that already shows this error gets its tools again at its next turn.
A Pi agent whose session file was moved or deleted starts a new session. Before, each turn stopped with "The Pi session file is not available on this computer."
A Pi, DeepSeek, Inception or OpenRouter agent with Workspace only access on macOS replies again. Before, each message stopped with "Credential store read failed" and "EPERM", or Pi stopped at the start.
Edit of a custom provider now shows the names of its saved headers. An empty value keeps the saved value, and you can change or remove one header. Before, the header rows opened empty, so the saved headers looked lost.
OpenBot no longer offers a CLI update for a provider that is turned off. When you turn the provider on again, the update is offered.
An answer through OpenCode stays whole when the model streams its reasoning between the words of the answer. Before, the answer was cut at each thought, so the start of a word moved to the steps and words showed in pieces, such as "B aca" for "Baca".
An SVG path in an agent answer was plain text. It is now a file link that opens the file.
An agent that tried to attach an SVG file got a general "not supported" error. It now gets told to write the file path in its answer, where the chat shows the image.
The working rings around an agent's face in the activity line of a chat now show in full. Before, the edge of the line cut them, most in a channel.
Go back from a thread to the agent's main chat at once. Before, the main chat was built again, which took about half a second and moved it to the top. Now it keeps its place.
Show the full time of each thread in the thread list. Before, a 12-hour time such as "10:12 AM" stopped after "10:1".
Close the side panel after you close its last tab. Before, the app stopped responding and the panel stayed open.
Keep the provider session of a thread when its delete fails and you restore it. Before, the restored thread started a new session and lost its thread instructions.
A message steered into a Claude turn just as the turn ends now shows its answer one time. Before, the answer showed again in the chat after OpenBot started again.
The browser client at openbot.run/app now shows the custom images of agents. Before, it showed the generated avatar for each agent with a custom image.
An agent with Full access no longer asks before each Computer Use action. Workspace only and Ask first agents still ask.
The usage limit notice in the composer now reads the ChatGPT account that the agent uses. An agent on a second account with quota left no longer shows the limit of the first account.
When you connect or resume Slack, Discord or Telegram on this computer, OpenBot now publishes this computer's server first. Before, the connection failed until you published the server or gave it a name in the server settings.
A Codex agent no longer says that it could not get an image after it generated the image. Before, the image showed in the chat, but the agent reported a failure and asked the user to try again. Each existing Codex chat starts a new Codex session one time after this update. To edit an image from before the update, attach it again.

Oct 26, 2026

Yaji 0.51.1

Fixed

On Linux, closing the OpenBot AppImage no longer shows a crash report for chrome_crashpad_handler. Before, its crash helper stopped with SIGBUS after each quit.
An agent on a shared server no longer tells you to open the chat on another device to use that device. The agent now says that it works only on the computer that runs the server. Settings › Computer Use also says that agents on another server use that server's computer.
On a Docker server of docker/instances with LXCFS, Gemini works again. Before, it stopped with SIGILL when it started, because the container listed only the CPUs of its limit as online.
On a server, Node programs that Claude agents start now also see the server's CPU limit. Before, only Codex and the ACP agents got it, and a build of a Claude agent started a worker for each CPU of the host.
DeepSeek, Inception and OpenRouter now download the Pi CLI when you save their API key. Before, the key was saved but the provider showed "The pi CLI is not installed" until you downloaded Pi from its own row in Settings.

Oct 26, 2026

Yaji 0.51.0

New

Choose the reasoning level of a new agent in the create-an-agent form. The model picker shows the level next to the model, as it does in the agent settings. The web app has the same choice.
Settings › General › "Open external links in" now lists each browser on your Mac, and "Other app…" lets you choose an app that is not in the list. On Linux and Windows, "Command…" takes a program path or command. Links in chats, sign-in pages, billing pages and connector pages open there. When that browser does not start, the page opens in the default browser and OpenBot shows a short notice.
Pin pages to the page rail. The new More button under the rail pages lists Routines, Workflows, Tasks, Schedule and Usage. Click a page to open it, or click the pin at the end of its row to show it in the rail or take it off. Routines, Tasks and Schedule start in the More menu. The choice is saved on this computer, and on the web app in this browser.
You can change the provider, account or custom agent while a turn or a task runs on this computer. The running work keeps its sessions. The new provider starts before the next turn after the running turn and its tasks end. The header shows the same "After this turn" badge that a model change already uses. On a remote server, a provider change still waits until the work ends.
Add DeepSeek as a provider. Connect it with a DeepSeek API key, then choose DeepSeek V4 Pro or DeepSeek V4.1 Flash for an agent. The agents run in the Pi CLI that OpenBot downloads, and the key stays encrypted on this computer.
A pasted image goes into the message box at the cursor, as a reference with a small preview, its name and its size. Thanks to @sebapaszynski.
When you copy or cut a draft that has attachments and paste it into another agent's message box, the files come with the text as new copies. The Copy action on a message also keeps its attachments. Other apps get the text with the file names. The clipboard holds no file paths and no file bytes.
The website has a Brand page at /brand. It has the OpenBot logo in each shape and color, with and without the name, as SVG and PNG files. "Download all" saves all of them in one zip file.
Right-click the OpenBot logo in the website header to copy the logo as SVG, download the brand files, or open the Brand page.
Agents use the skills in your ~/.agents/skills folder with every provider. Claude gets them as agents:<name>, and a skill that Claude already reads from ~/.claude/skills is not added again. Antigravity, Muse and custom ACP agents get a list of these skills in their instructions and read a skill's SKILL.md when a task needs it. The other providers already read the folder.
Type $ in the message field to tag a skill from ~/.agents/skills. The agent gets the path of the skill's SKILL.md, so every provider can open it.
You can now zoom into an image in the image viewer of a chat. Pinch, or hold Ctrl or Command and scroll, to zoom around the pointer; double-click to zoom in or back to the whole image; drag or scroll to move a zoomed image. The viewer also has zoom out, zoom in and fit buttons, and the +, - and 0 keys. Before, you had to download a large image to see its detail.
Add Apple apps to the Apps catalog. On a Mac, agents can use Mail, Calendar, Contacts, Reminders, Notes, Messages and Safari through the community server mcp-icloud. It needs no account or key; macOS asks for Automation access the first time an agent uses each app.
Add OpenRouter as a provider. Connect it with an OpenRouter API key, then choose one of the models that OpenRouter serves for an agent. The model picker groups the models by lab and marks the free models. The agents run in the Pi CLI that OpenBot downloads, and the key stays encrypted on this computer.
Add Supabase to the Apps catalog. After you sign in to Supabase in the browser, agents can list projects, run SQL, read table schemas and logs, check advisors, manage edge functions and branches, and search the Supabase docs. Agents get full access, so they can also change data, apply migrations and create or pause projects.

Improved

Claude agents use Claude Code's own prompt cache time again by default: 5 minutes with an Anthropic API key, and 1 hour for the main chat of a Claude plan within its usage limits. In 0.50.0, OpenBot asked for 1 hour for all agents, and a 1-hour cache write costs more. The "Prompt cache" setting on the Claude page in Server settings › Providers now has "Claude Code default" and "1 hour". If you chose "1 hour", it stays.
The website FAQ answer to "Is OpenBot open source?" now uses simpler words. It says that all of the source code is public, links to it on GitHub, and says that it is free for noncommercial use. The launch news post now calls OpenBot source-available, not open-source.
The website header has a Resources menu with Download, Changelog, Community, Documentation, Brand and Status. Changelog and the Community button moved into it.
When OpenBot updates to a new version, it also updates the provider CLIs that it installed, such as Claude, OpenCode and Grok, with no update notification. A notification shows only when an update fails, with a Retry. A CLI that you installed yourself is not changed.

Fixed

A server with a saved Anthropic API key no longer starts a Claude sign-in on the API key account, or removes that account, when an older OpenBot version asks for it. That version shows the key as a signed-out account. Now it gets a message that tells you to change the key in Server settings › Providers.
Open and create agent routines in the web app. Before, Create routine and a routine in Schedule showed the "This page could not load" error page.
Show why an OpenCode agent does not answer when the free models reach their usage limit. Before, each message failed with no error, and only the agent's face changed.
Pi agents no longer stop with "Pi returned an invalid RPC record." when a tool call has a value such as true that is also the value of an MCP server setting or of the GitHub connection. Before, OpenBot stopped Pi and started it again, and the turn failed.
The Cursor usage ring no longer shows an "Other models" window for a Cursor plan that includes no other models, such as an X subscription. It shows the plan's total usage only. Before, it showed an "Other models" window that the plan does not have.
Escape in the model picker of the create-an-agent form now closes only the picker. Before, it also closed the form and the draft was lost.
In the web app, browser Back and Forward, including the side buttons of a mouse, now open the chats that you opened before, also on another server. Before, Back left the app and opened the page that you came from.
On macOS 26 with the Dark icon style, the OpenBot app icon is a black tile with white eyes. Before, macOS made a black tile with black eyes, and the eyes were almost invisible. The Light style still shows the lavender icon.
The "Open external links in" setting had no effect, and it went back to "Default browser" after a restart. Now OpenBot keeps it, and "OpenBot" opens links in chats in the built-in browser.
The Linux AppImage starts on Arch Linux, Omarchy and other systems with no libfuse2. It mounts itself with fusermount3 from the fuse3 package. Before, it stopped with an error about libfuse.so.2.
In a wide window, the agent's working row and its pending questions, approvals and other cards now show in the centered chat column. Before, they showed at the left edge of the pane until the turn finished. (#3146)
Give MCP server tools to Codex agents on GPT-6 and other code-mode models after an update to Codex 0.164. Before, a chat that started before OpenBot 0.50.0 had no MCP tools. Each Codex agent starts one new Codex session to get the tools.
On Windows, OpenBot now tries again when another program, such as antivirus, has a settings file open while OpenBot saves it. Before, a save at sign-in could fail, and then Slack, Discord and Telegram connections failed with "Sign in, give this computer a name, and try again."
On Windows, Claude starts when the Claude Code CLI comes from npm (claude.cmd). Before, every Claude session, model list and usage read failed with "spawn EINVAL", and sign-in and sign-out through such a CLI failed too. Stopping a session now also stops the Claude process that the .cmd file started.
Keep the web client connected while a server sends many events. Before, it disconnected and connected again every few seconds during a turn.
OpenBot no longer shows "OpenBot database is not initialized." when you quit it. The windows now hide when the quit starts, and the database waits a short time for a slow agent stop before it closes.
A sleeping cloud server keeps its data when the hosting provider refuses its keep-alive start or its start for a routine run. Before, a refusal such as a billing error left the server in an error state, no automatic start tried again, and the provider deleted the server's disk 14 days after its last stop.
Continue unfinished agent work after a provider reaches its usage limit and switches to a backup. Keep completed work, tool results and attachments with the request, including across backup models.
Deleting an agent that runs on its own provider process, such as an agent with "Workspace only" access or a second provider account, now stops that process first. Before, on Windows the delete could fail and the agent came back until you restarted OpenBot.
When a delete fails after the agent is already gone, the agent no longer comes back in the list until a restart.
In a channel, an agent that posts its task result no longer shows a second short reply, such as "Yes.", under it. Before, the text that the agent wrote after the result showed as a separate message. Now it shows in the steps of the turn.
The agent's browser evaluate tool no longer times out on pages with an ad frame that does not run its timers, such as Amazon product pages.
On a server, agent builds no longer take all the CPU and memory and drop your connection. Before, a build such as next build started one worker for each CPU of the host, not of the server's limit, the server stopped answering, and the apps reconnected every few minutes. Agent processes now run at a lower CPU priority, and Node programs of agents see the server's CPU limit.
When a server is almost out of memory, OpenBot now stops leftover agent processes first, such as preview servers and build workers that stayed after their job ended. Before, they kept their memory until the kernel stopped a process or the server restarted.
Routine wake-up commands now run also when the long command fails, so an agent learns that its job stopped.
OpenBot now starts the Codex app server again when it stops answering. Before, a hung Codex process stayed in use, and every agent turn and routine run failed until you stopped codex.exe yourself.
When Codex does not answer, a message now fails with that reason. Before, it failed with "OpenBot could not register MCP servers in the Codex configuration", although config.toml was correct.
Usage now shows the token cost of OpenCode turns, including OpenCode Go. Before, the OpenCode cost was always "Unavailable". The cost is the session cost that OpenCode reports, at API-equivalent rates, not the subscription charge. When earlier turns of a session have no cost, as in a session that started before this release, the first turn with a cost has no cost.
Two reads of the history of one Claude session at the same time no longer fail with "database is locked". Before, this error could show when an agent continued on its backup model while OpenBot read the session history after a start.
Connecting Firecrawl, or another plugin that starts with npx, no longer fails with "The server did not answer in 10 seconds" on its first start. A local server now has 60 seconds to download its package and answer. A server at a web address still has 10 seconds.
When you drag a channel to a new place in a sidebar section, it now stays where the sidebar showed it would go. Before, the first drag after you moved a channel or an agent into the section with "Move to", or made a new one, could put the agents above the channels or put the channel in a different place. On a team server, the server must also have this version.
The model picker for a new agent no longer starts on Claude when Claude is not installed or not signed in. It now starts on a provider that can run the agent, as the server already does.
When an agent runs on a custom endpoint and the endpoint does not answer, the waiting notice and the error now name your endpoint. Before, they said "OpenCode has sent no update" and "OpenCode stopped responding", so it looked as if OpenCode's own models were in use.
A message that you steer into a running DeepSeek or Pi agent now goes into the running turn also while Pi compacts its context. When Pi refuses a steer, only the steer fails: the agent keeps working, and the message waits in the queue. Before, the refusal stopped the turn in the chat while Pi kept working, and the next messages failed with "Agent is already processing".
A "Could not steer the agent" error no longer stays below the last reply after the turn ends. Before, it stayed at the end of the chat until you sent the next message, so it looked like a new message.
When you sign in a provider account of a joined server in this computer's browser, the dialog no longer says "This sign-in ended" while the server completes the sign-in. After the browser returns, the dialog shows the progress until the server gives the result.
A channel whose agent waits for your answer, your approval, or you in the browser now moves to "Needs you" at the top of the sidebar, with the same badge and tooltip as an agent. Before, only the channel conversation showed it, so a request in a channel was easy to miss.
Gemini (Antigravity) agents in a channel get only the messages they have not read. Before, each channel turn sent the complete channel history again into the same session, so usage grew with each message. Long channel sessions are now also compacted.
A change to MCP servers or skills keeps the conversations of Claude agents. Before, each Claude agent started a new session on its next turn, with a summary of the history and no prompt cache.
Agents in a channel report their result with fewer model requests. Before, an agent such as Grok searched for the channel tools first, and each search step sent the whole conversation again.
An "always allow" approval for an MCP tool keeps the conversations of Codex agents. Before, each Codex agent that used that server started a new session on its next turn.

Oct 25, 2026

Yaji 0.50.0

Action needed after you upgrade

A sign-in to a Google MCP server, such as Gmail, Calendar or Drive, now keeps working after the first hour. OpenBot asks Google for a refresh token. Sign in to these servers again after you upgrade.

New

Show a storage card at the bottom of the sidebar when the disk of the selected server is almost full. A ring shows the used part of the disk. "Free up space" opens Server settings › Storage. You can close the card while the disk is almost full; it comes back, in red, when the disk is full. Only an owner or admin of the server sees it.
"Open in new window" in the menu of an agent in the sidebar opens its chat in a window of its own, for example on a second display. Both windows show the same live chat, and a window mark beside the agent's name shows which chats have a window. Closing the window keeps the chat in the main window.
A message that an agent sends to or gets from another agent now opens in the chat. Click the arrow on its line to read the text, and the message it answers.
You can add more than one Google account to Gemini. Select Add account on the Gemini row, as for Codex, Claude and Grok, and choose the account for each agent. Each account keeps its own sign-in and usage. A saved Gemini API key applies only to the first account. This needs Gemini runtime 1.4.0 or later. On a joined server, the sign-in page opens in the server's browser, and you use it through the live view.
Grok agents can now be steered. A message that you send in Steer mode while a Grok agent works goes into the running turn, and Grok reads it after its current step. Before, the message waited in the queue until the turn ended. A Grok CLI without this feature keeps the message in the queue.
Settings › Agents sets the defaults of the agents on this computer: Local scripts, Agent starts new chats, Show the plan before long tasks, and Steer agents while they work. Each setting shows which agents have it on. Its page changes the default, the value of one agent, or the value of all agents with Set all.
An agent that has no value of its own follows the default. The defaults start off, so no agent changes until you change a default.
Give one server a custom member limit, such as for an enterprise customer. The limit replaces the limit of the server's plan.
The Claude page in Server settings › Providers has a "Prompt cache" setting: 1 hour or 5 minutes. An owner or admin can set it for this computer and for a joined server.
Korean interface. Choose 한국어 in Settings > General > Language, or keep System default on a computer that uses Korean. The web client uses it too.
A custom ACP agent can have a usage command, such as /usage. OpenBot sends it at most every 5 minutes, and the agent's logo in the rail shows the last answer when you hover it. Use a command that the agent answers itself: other text starts a model turn each time.
Server settings has a Vault page, after General. Save logins, payment cards and contact info on this computer, encrypted by the operating system. Agents sign in with the saved logins, fill a saved card on a checkout page, also in the card frames of payment providers, and type the saved contact info. An agent sees only the card name and its last four digits, never a password, a card number or a security code. Import logins from a password export in CSV format. The Vault is on the local server only.
The first time an agent fills a saved card on a site, the server owner approves it. Later fills on that site need no approval. Remove a site from the card in the Vault.
Agents can add and change contact info in the Vault, change the username of a login and the name and expiry of a card, and delete items. The chat shows each change, and the Vault shows which agent added an item. On the secure password card, Save to Vault keeps the password that you type.
openbot.run has provider pages for Muse Code, Pi, Inception and ACP agents such as Goose, Qwen Code and GitHub Copilot CLI.
The steps row above an agent's answer now opens a trace of the turn. Each thought, tool call and subagent is a bar on one timeline, with the duration, cost and tokens of the turn. You can search the steps with ⌘F (Ctrl+F), fold an agent's steps, replay the turn at up to 10× speed, and move between the rows with the arrow keys. A click on a step shows it in the chat.
The steps row now looks like the subagents header above it. It shows the time of the turn and how many steps failed.
Simplified Chinese for the desktop app and the web client. Select 简体中文 in Settings, or use a Chinese system language.
The usage list shows each Gemini limit on its own line, such as Gemini Models, Gemini 4 Argon, and Claude and GPT models. The warning above the message box uses the limit of the agent's model.
Claude can use an Anthropic API key. Open the Claude page in Server settings → Providers and select Add key. The key shows as an "API key" account. To bill an agent to the key, select the API key account in the agent's model menu. Other agents keep the Claude plan. When Claude has no sign-in, all agents use the key. A joined server shows the key when the server also has this version.
Add Cloudflare to the Apps catalog. After you sign in to your Cloudflare account and choose Read only (or more) on its consent page, agents can work with Workers, DNS, R2, D1 and KV, query Workers logs and search the docs.
A Workflows page in the page rail. Draw a workflow on a canvas with agent, model, condition, approval and output steps, and connect them. The workflow saves while you edit it.
Run a workflow, or one step of it. You see each step's status and answer on the canvas, and you can stop, retry or approve a step.
A condition step asks a model a yes or no question. Only the steps on that branch run.
A workflow can loop: connect a step back to an earlier step, and the steps between them run again with the new answer. The canvas marks each loop, and warns when no condition can end it.
An output step shows a notification or posts the result to a connected messaging place. Only an owner or admin can choose the place.
A workflow can start on a schedule or from a signed webhook request.
Version history: preview an earlier version of a workflow and restore it. Undo gets back the canvas from before.
Members of a joined server can open and run the workflows whose agents they can use. To change, delete or restore a workflow, a member needs “Can manage” on each of its agents; otherwise the canvas is read-only.
Usage shows the tokens of workflow model and condition steps in one Workflows row. Steps that use an agent count under that agent.
Consecutive model and condition steps reuse one started provider for 5 minutes, so they start faster. Each step still starts with no earlier text.
When analytics is on, OpenBot counts workflow actions and runs: the kind of action, the outcome, the duration and the step counts. It never sends a workflow name, step text or graph.

Improved

The Linux docs and the Linux download page on openbot.run now name Ubuntu 26.04 LTS. OpenBot 0.48.1 runs on Ubuntu 26.04 with the Electron sandbox on.
When nothing comes from the provider for a minute during a turn, the activity line says "Waiting for {provider}…" and stops its shine. A notice above the message box says how long the provider has sent no update. Before, the line looked busy the whole time, also with no internet.
When the computer has no network during a turn and nothing comes from the provider for 5 seconds, a "No internet connection" notice shows above the message box, and the activity line says that it waits.
On a joined server, Add account for Codex and Claude opens the sign-in page in the browser of your computer, as the first account does. Before, the page opened in the server's browser, and you used it through the live view. That was slow, and a magic link or a CAPTCHA could stop the sign-in. A server on an older version still uses the live view.
On a very wide window, such as an ultrawide monitor, the chat now stops at a 1440 px column in the center of the pane. Before, the user's messages were at the right edge of the pane and the agent's replies at the left edge. A laptop window shows the chat at full width, as before.
When you try to turn off a provider that an agent uses, a toast now tells you why the switch stays on. Before, the row showed red text with the names of all the agents. When more than one agent uses the provider, the toast gives their number.
"Steer agents while they work" moved from Settings › General to Settings › Agents.
The sidebar no longer shows a presence dot on a person's avatar. The dot sat outside the shape of avatars with no photo. The conversation header still shows whether the person is online.
The app demo on the home page now shows a team of four agents, each on a different provider: Claude, Gemini, ChatGPT and Grok. The lead agent plans a launch, sends work to its teammates, asks you a question and answers with a table and a diagram. Each teammate plays its own short scene when you open it, and a message that you type gets a reply. The sidebar groups the chats in sections and shows a launch channel with a chat history.
Claude agents now keep the prompt cache for 1 hour by default, also with an Anthropic API key, for subagents, and on a subscription that uses extra usage. Before, these used 5 minutes, so a turn that started later sent the full context again at the full price. A 1-hour cache write costs more than a 5-minute write; choose "5 minutes" to go back to Claude Code's own choice. A promptCacheTtl in the Claude settings, or CLAUDE_CODE_PROMPT_CACHE_TTL, still sets the time.
In a channel, the activity row now shows the name of each agent that works, before the activity text and in the colour of the agent's messages. Before, it showed only their faces.
Agents now follow your user-level CLAUDE.md or AGENTS.md, and the one at the root of the agent workspace, when these files set the persona, reply length and format, delegation, or group-chat behavior. Before, OpenBot's built-in style guidance and your files gave the agent two rule sets for the same decision. Files in cloned repositories do not get this priority. The rules for tools, memory, approvals, secrets, private details, safety checks and the browser do not change.
A focused field shows a grey border, and a focused button shows a thin grey ring. They no longer show a blue glow.
The comparison pages on openbot.run now list all the providers that OpenBot runs, including Muse Code, Pi, Inception and ACP agents, and 11 app languages. New rows compare backup models and logins and payments, and most pages also compare voice and Slack, Discord and Telegram.
The home page and llms.txt on openbot.run now name the new providers and backup models.

Fixed

An agent set to switch to a backup model without asking now switches when its provider is overloaded or down. Before, it showed the error and stayed on the failing model for Claude's "529 Overloaded" errors, Codex's server, overload and retry-limit errors, Gemini's UNAVAILABLE status, and dropped connections.
A channel message that stopped because the channel had no lead now continues when you choose a lead. The new lead routes it. Before, it stayed stopped until you clicked Continue.
Channel members now know which member is the lead.
The x64 remote desktop runtime now starts on Ubuntu 26.04 LTS. Before, it needed libminiupnpc.so.17 and libicuuc.so.74, which Ubuntu 26.04 does not have, and the Docker image with BASE=ubuntu:26.04 did not build.
Remote desktop now starts in the Docker image on Ubuntu 24.04. Before, Sunshine stopped because libnuma1 was not installed. The server install now also installs it.
When a Grok turn fails after it used tokens, the error now shows the reason that Grok gives, such as empty response from model (reasoning_only), and not only Internal error. The routine run history keeps the same reason (#2980).
When you stop a sign-in that a joined server relays to your computer, the provider list now shows the change at once. Before, a new account that you stopped stayed in the list for some seconds.
An app that an agent suggests in a chat, such as Linear, now shows a green Connected button with a check after you connect it. Before, the card kept its Connect button.
When an agent asks another agent for help, the "Waiting for replies" list now shows when that agent waits for your approval, your answer, the browser, or a usage limit. A button opens its chat, where you can answer. Before, the list showed "Working" until the wait ended.
The agent usage page now counts all Gemini tokens. Antigravity 1.4.3 sends the tokens of each turn, but OpenBot added them as a session total, so a turn smaller than an earlier turn in the same session counted as 0. Gemini thinking tokens now also count as output. Usage that was recorded before this fix does not change.
The app demo on the home page no longer shows a time zone prompt, an update card or a helper agent from another server. Its dates now show today.
On a phone, the app demo on the home page now shows all four agents in one row and the chat under them. Before, only one agent showed and the chat area was empty.
Support service usernames with uppercase characters in server setup and update scripts.
Gemini now shows the Claude and GPT models that your Google plan includes, such as Claude Sonnet and Claude Opus. Before, the model menu showed only Gemini models.
Claude agents use less of the Claude plan. A sidebar move or a teammate rename no longer starts the Claude conversation again with a new system prompt: the list of section teammates now goes with the next message, as the saved memories do. Before, the full conversation was sent to the model again and written to the prompt cache.
A Claude conversation no longer restarts when it is opened again with the same settings after a model or effort change.
Claude, ChatGPT and Muse channel members use less of their plan. A member's next assignment in the same conversation gives only the channel messages that it has not read. Before, each assignment sent the recent channel history again, up to 60,000 characters. The first assignment after a start of OpenBot, a new conversation, a provider change, or a compaction still sends the complete history.
A provider account that you add now shows its usage rings in the page rail when its sign-in ends. Before, the rings showed only after a reload or the next usage refresh, up to 5 minutes later.
On a joined server, the usage rings no longer go away when the host resets the usage of a provider account.
When two or more notifications were open, you could not reach a notification under the first one. The stack closed when the pointer moved onto it, because the notifications were over the part of the window that moves the window. Only the buttons of the first notification responded.
The waiting block, the queue, a reply and the notices above the message box are now one card, with a line between each part. Before, each part was a separate card, with gaps and cut edges between them.
An agent can click a choice, such as a payment method, when a decorative layer of its label covers the radio button.
The first Claude turn after OpenBot starts again now has a cost in Analytics and in the turn trace. OpenBot estimates it from Anthropic's list prices, because Claude does not report the cost of that turn.
The Live Activity updates that OpenBot sends to your iPhone cut a long code span with emoji at the correct length. Before, the cut could split an emoji into half a character, and an emoji counted as two characters, so the text stopped too early.
The Live Activity updates that OpenBot sends to your iPhone show a reply that starts with inline code in three backticks, such as ``` `npm install` first ```. Before, the update read that line as the start of a code block: the line disappeared and the next lines showed as code.
Gemini usage now follows the limit of the model that an agent uses. Google can give a model, such as Gemini 4 Argon, a weekly limit of its own. Before, OpenBot put all Gemini limits together, so a spent Argon limit could show for Gemini 3.8 Flash, or Argon could show room that it did not have.
Stop showing a "Provider error" when Codex completes a credential storage fallback. Keep the redacted diagnostic in the provider log. Authentication and storage failures still report errors.
A custom provider with a base URL on api.anthropic.com now uses Anthropic's Messages API, so OpenCode adds prompt cache breakpoints. Before, it used Anthropic's OpenAI-compatible endpoint, which does not cache, so each turn paid the full input price.
Usage and cost for OpenCode turns now count the uncached input, the cache reads and writes, and the thinking tokens of each turn. Before, a turn could show no uncached input.
Analytics now records why a provider download failed, for example no disk space or a failed checksum. Before, each failure was recorded as unknown.
When you save or remove a provider API key, agents that a usage limit held now continue at once. Before, they waited for the old plan's limit to reset, which could take days.
Sign in to an MCP server again when OpenBot cannot read the saved sign-in file. Before, Connect stopped with "The MCP sign-in file is unreadable" until the file was deleted, which dropped every saved MCP session.
Restart now closes turn rows that stayed running although no thread names them as its active turn. Before, such rows stayed running for good, as the follow-up data on #2214 showed.
An OpenCode agent with more than one chat keeps its OpenBot tools. Before, about 10 minutes after its other chat went idle, every OpenBot tool call failed with {"error":"unauthorized"}, including routines, teammates and the browser, until OpenCode restarted. Each tool call also runs in the chat that made it, when two chats of the agent work at the same time.
A long channel keeps working when its history summary comes back too long. OpenBot now asks the lead for a shorter summary, and asks once more when the result is still too long. Before, every new assignment in that channel failed with "The history summary is invalid" and "The provider could not start this assignment".

Oct 24, 2026

Yaji 0.49.0

Action needed after you upgrade

On macOS, sign in to Gemini again after you upgrade.

New

You can change the key of the server on this computer in Server settings › General › Server key. Computers with this version or later, the web client and webhook routes stay connected. Before, a new key showed "Server identity changed" on each device, and each phone had to scan a new code. Older versions of OpenBot and the current phone app still show "Server identity changed". Do this if the key file was copied or exposed.
The website compares OpenBot with Viktor at /compare/viktor: models and price, where agents work, what each one reads, who holds your data, teams, approvals and integrations.
Connect Bitwarden or a self-hosted Vaultwarden on a joined server. In Server settings › Connectors, an owner or admin enters the server address, the personal API key and the master password. The server signs in with its own Bitwarden CLI, downloads it when it has none, and unlocks again by itself after a restart. Agents on that server can then sign in to sites with the logins in the folder Shared with OpenBot. This works from the desktop app and the browser client.
The website compares OpenBot with Paperclip at /compare/paperclip: setup, phones and remote access, how you work with agents, price, and license.
When a task needs an API key or a token, the agent now shows a secure form in the chat, as it does for passwords. The key is saved for that agent on the computer or server that runs OpenBot, outside its workspace, and the agent reads it from a file. It does not go into the chat, the logs or a tool argument, and you do not need to connect to the server to give it. Only the server owner or an admin can enter a key. To remove a key, ask the agent to remove it. Deleting the agent deletes its keys.
Mobile Connect settings can keep chats and files on a private network. Choose "Private network", and phones, other computers and team members connect to this computer only over Tailscale or the same Wi-Fi. They still sign in with OpenBot, but chats and files never go through OpenBot servers. Settings shows when Tailscale is connected.
Let an owner or admin of a connected server, self-hosted servers included, change "Tasks at once" for each agent of that server in the agent's Advanced settings. The server saves and applies the limit. Before, only agents on this computer had the setting.
While an agent works, a chevron on its activity line opens the steps of the turn: its thinking, its tool steps, its memory searches and the skills it uses.
A finished answer has a closed "N steps" row above it with the same steps, and the memories that the turn saved, updated or forgot. Tool steps stay until the app closes; thinking and memory changes stay after a restart.
The "Show agent steps" setting (Settings > General, and Preferences in the web client) opens the steps by default on that device.
A light theme on the desktop and in the web client. Settings > General > Theme has System, Light and Dark. Dark stays the default, so the app does not change after the update. In the web client, the choice is in Preferences and this browser keeps it.
On macOS, Settings > General has a Transparent window switch under Theme. When it is on, the main window shows the blurred desktop behind its frame and pages. Dialogs, menus and cards stay solid. While it is on, visited websites and the native menus use the app theme, not the computer's setting.
A visual reply has a full-screen button. The page keeps its state when it goes full screen and back. Esc or a click outside the page closes the full-screen view.
A click on an image in a visual reply opens the image in the image viewer.
Search, Settings and the server number shortcuts work while a visual reply has the focus.
A visual reply can answer the agent. A button in the page puts text in the message box, and you send it. The page cannot send a message itself.
A visual reply can give the agent its current state. A button in the page puts the state in the message box as a quote of the page, and your next message sends it.
A new chat with an agent is no longer blank. It shows the agent, its description, and a faint sample turn: the agent's own routine is asked, and the steps of the turn fill in one at a time. With reduced motion, the sample turn stands still.
Named browser profiles. In agent settings > Advanced, an agent can use the shared profile or a named profile. Agents in one named profile share its sign-ins, so two agents can work in one account while other agents use other accounts on the same site.
"New profile" at the end of the browser profile list makes a named profile and gives it to the agent in one step. Server settings > Browser profiles shows the shared profile and adds, renames, clears and deletes the named profiles. "Manage" in agent settings opens it. A deleted profile's agents change to the shared profile.
The browser toolbar shows the profile of a tab that is not in the shared profile.
On a joined server, an owner or admin manages the server's browser profiles and chooses each agent's profile, on desktop and on the web. Each member sees the profile of the server's tabs. This needs no setup: the controls show when the server runs this version or later.
Each named profile has a color, and the shared profile is gray. The browser profile list, the browser toolbar and a line under each tab show it, so the tabs of different accounts look different. A new profile gets a color that no other profile uses, if one is free. Server settings > Browser profiles changes it.
The Tasks page before the first task is no longer blank. It shows a short pitch, the composer under it, and a faint sample list: a task arrives in Waiting, moves to Working, then to Needs you as Answered, and fades. With reduced motion, the list stands still.
Back and Forward arrows in the sidebar top bar open the chats that you opened before, as in a browser. The keys are ⌘[ and ⌘] on macOS and Alt+← and Alt+→ on Windows and Linux. The side buttons of a mouse also work. The history stays in memory for the open server.
On macOS and Windows, the "Keep this computer awake" setting (Settings > General) stops the idle sleep of the computer while it is connected to power, so routines run on time. It is off by default. On battery, the computer sleeps as usual, unless you also turn on "Also on battery". The display can still turn off, and a closed laptop lid or a sleep that you select still sleeps the computer.

Improved

Each plugin page on openbot.run now shows what an agent can do with the plugin, how to connect it to OpenBot, common questions, and the sources for each statement.
The guides "OpenBot 101" and "WTF Is OpenBot?" now have the titles "How to Use OpenBot: Set Up Your First AI Agent" and "What Is OpenBot? Models, Providers and Agents Explained". Their addresses did not change.
A message that you send no longer shows "Sending…" or turns grey. Only a failed send shows: "Not sent" and its reason beside the time, with Retry, Edit and Dismiss buttons.
Product analytics on turns sends how the turn is paid for: a plan, an API key, a free model, or your own endpoint. It does not send keys or accounts. The privacy policy lists the new data.
The web client and the desktop app on Windows and Linux show the OpenBot logo in the top-left corner of the window. Before, this corner was empty, because only macOS shows window buttons there. In the web client, the logo opens the OpenBot site.
On these windows, the server menu starts at the same left edge as the search field and the sidebar rows.
A schedule with no routines now shows a faint month where sample runs land and write a short note of what they did, with "Set it once. Wake up to it done." The "Create Routine" button asks which agent to use when the server has more than one, and opens that agent's routine settings on a new routine.
An agent or a channel with no memories now shows a faint memory list that fills in by itself: each sample memory writes itself in handwriting, then becomes a normal row. The title says "Tell it once. It will remember.", and an "Add memory" button opens the editor.
An empty "Pending invitations" list in server settings now shows faint sample invitations that appear one at a time, and an "Invite people" button that opens the email invite field. Before, it showed only "No pending invitations."
An agent's routine canvas with no routine shows how a routine works. A faint sample diagram builds itself, and one run goes through it. "Create Routine" opens the agent's settings on a new routine. With reduced motion, the canvas shows the finished run and does not move.
The "Own browser session" switch in agent settings is removed. An agent that uses its own session keeps it, and the browser profile list shows it as "Only this agent" until you choose another profile. Use a named profile for a new agent.

Fixed

An hourly routine can now run between two times that cross midnight, for example from 8:00 PM to 10:00 AM. Before, the routine form showed "End time must be after start time" and did not save the routine. A run after midnight uses the days of its own date: with Monday to Friday, a run at 2:00 AM on Saturday does not occur.
A long chat keeps the older messages that you scrolled back to when a new message, a routine update or a reaction arrives. Before, the chat showed only its latest 100 messages again, and a message that you opened from search jumped to the end of the chat.
Fold the thinking of a channel member into steps, as in an agent chat: under the activity line while the member works, and above the member's answer after. The unread count of a channel no longer counts this thinking. Before, each part of the thinking showed as a message, and a model that sends all its reasoning, such as DeepSeek, showed a very long message.
Keep a failed Codex tool call, such as an image generation that fails, out of the ChatGPT provider error notice. Before, the notice stayed on screen in each chat that you opened, also in chats of other providers. The agent still gets the error as the tool's result.
After an in-app update installs, OpenBot deletes the downloaded update file when the new version starts. Before, a copy of the full update stayed in the updater cache on macOS and Windows. A downloaded update that is newer than the running version stays until you restart.
A long preview in an agent's confirm card now has a "Show all" button that opens the full text, and "Show less" closes it again. Before, the preview stopped at a fixed height with no visible sign that more text was there.
On macOS, a Gemini agent with Workspace only access now uses the Google sign-in. Before, it stopped with "Gemini did not return an authenticated account" while Settings showed Gemini as connected.
The web client connects to a server again in seconds after a connection stops working. Before, on a server that uses a relay, it could wait up to a minute when it still received data from the server.
In a channel, Open on a "Complete the step" browser card now opens the page in the chat of the agent that asked. Before, Open did nothing, and the card showed "Browser page" with no preview.

Oct 23, 2026

Yaji 0.48.2

Action needed after you upgrade

On Windows, Restart and install updates OpenBot with no installer window, then opens it again. The update from 0.48.1 or older still shows the installer window once. Select Finish to open OpenBot again. Your settings, agents and chats stay.

New

In Server settings > Roles, a role and Everyone have a Set all menu. It gives one access level to all agents at once, so you do not set each agent one by one.
Voice prompts and read-aloud replies work on Linux (x64 and arm64), as on macOS and Windows. Before, the Linux app showed no microphone and no Voice settings.
You can now duplicate a whole sidebar section from its menu, in the desktop app and the web client. OpenBot makes a "copy" section right after the first one and copies every agent in it. Each copy keeps its settings and files, and starts with no messages. If one agent cannot be copied, for example because it is busy, the copy stops and keeps the agents it already made.
"Send test event" signs an event that matches the routine and receives it as a real request. The routine runs once, and the editor says whether the run started or why it did not. A joined server must also run this version.
The Marketplace has the Dr. Eggbot agent. It designs focused agents, then creates them and checks the result. It installs with pstack, five skills of its own, and two paused healthcheck routines: a weekday scan of agent chats for friction and a Monday audit of routines for token waste.
An agent can now give several agents the same provider, model, reasoning effort, access, Computer Use, or notifications setting in one step, for example when you ask it to move your team to a new model. It tells you the result for each agent, and a failure for one agent does not stop the others. As before, an agent can only restrict access and Computer Use.
Agents can see the accounts of each provider, and read and set which account an agent uses and its backup models, with their accounts, when you ask them. Agents see the account names, never the credentials.
When an agent's backup models are set to Automatic, a provider error now moves the agent to the next backup model too. This covers a spent balance, an unavailable provider, a rate limit, a sign-in error and a network error. The message runs again on the backup when the agent had not started to write, and the agent goes back to its own model after 10 minutes. Before, only a usage limit moved the agent.
read_agent now says when a provider error, not a usage limit, keeps an agent on a backup model.
Show the Cursor plan usage next to the other providers' usage. The Cursor ring shows the usage of Cursor's own models and of all other models for the billing cycle. OpenBot reads the Cursor CLI's sign-in to ask Cursor for this usage.
Right-click a message to reply to it, copy it, react to it or read it aloud. React opens all reactions in a submenu. Shift+F10 opens the menu from the keyboard. Over selected text or a link, right-click still opens the menu that copies them.
An agent can now mark its routine run as failed when the task could not be done, for example when a server it checks is down. The run history shows the run as failed, with the reason. The answer still shows in the chat and still goes to the routine's result destination. In a routine flow, the agents after a failed step are skipped.
On Windows, the installer lets you choose the disk and folder for OpenBot. Before, it always installed to %LOCALAPPDATA%\Programs\OpenBot. Updates keep the folder that you chose.
A channel member can now change its channel when a person asks for it in the channel: add or remove a member, change the channel instructions, and make, change or delete the channel's routines. Each change shows in the channel with the agent and the person who asked. A routine run or an agent's own request cannot make these changes.
Agent settings have a new switch, "Own browser session". When it is on, the agent's browser tabs keep their own cookies, sign-ins and site data, apart from other agents. Two agents can then use different accounts on the same site, such as WhatsApp Web or Gmail. The switch is off by default, so agents keep the shared session and its sign-ins. A change closes the agent's open tabs. Deleting an agent deletes its own browser session.
Reset the usage limits of a Codex account from Settings > Providers. When Codex gave the account a usage reset credit, the account's ⋯ menu shows Reset usage. OpenBot asks for confirmation, spends one credit, shows the result, and starts the agents that the limit held. This works for each Codex account, also on a joined server where you are an owner or admin.
Sign in to a site for a server's browser on this computer. In the browser's page menu, select "Sign in on this computer" and sign in in the window that opens. When the sign-in is complete, OpenBot sends only that site's sign-in to the server, the tab loads again signed in, and the window closes. If the window stays open, select Done. This helps with sites such as X that refuse a sign-in from a server address. The server must run this version.
When a site refuses a sign-in from the server, the agent shows a card that says so. In the desktop app, the card has one main button, "Sign in on this computer". After the sign-in, the card shows that you are signed in, and the agent continues.
The website has a privacy policy at openbot.run/privacy-policy. The "Privacy policy" link in the site footer opens it. Before, the footer linked to the privacy notes on GitHub.

Improved

When you open an agent, the cursor goes into its message box, so you can type immediately. Before, the focus stayed on the item that you clicked.
An agent needs fewer steps before it uses a tool, so a run, such as a routine, sends fewer input tokens. A tool search now also gives the schema and guidance of its best matches.
When you type @ and part of a name, the agents whose name matches show first, then the agents whose title matches, then the agents whose description matches. The match ignores accents, so e finds é. Before, the list kept its order for all matches.
OpenCode answers now appear while the model writes them, not only when the answer is complete.
In the desktop app, the error above the message box now names the provider and model of the run that failed, so you can report it.
A webhook routine now shows its setup as three steps: copy the endpoint, add the signing secret to the sender, and send a test event. A new routine has a "Save and get the endpoint" button, and "How to sign a request" shows a curl example. Schedule and Webhook are now two buttons above the trigger, not an item in a menu.
The agent and channel settings panel opens wider, so routine settings have more room.
Save now tells you what is missing, under the name, the instruction or the filter, instead of staying unavailable.
The Starter plan of hosted servers is sold out. The add-server dialog shows it as sold out, and a new server must use the Standard or Pro plan. Servers that have the Starter plan keep it.
A server with 5 GiB of memory or less, such as the Starter plan, now runs at most 2 turns at the same time. Other messages wait in the queue for a free slot. Before, 4 turns could start together and show a low-memory notice on a server that still had memory.
Voice recording warns you before its 2-minute limit. The last 15 seconds count down in the recording bar, two short pips play at 10 seconds left, and a chime plays when it stops. Before, the recording stopped without a sound and you could keep speaking without knowing. A notice above the message box now says why the recording stopped.
The model picker now shows only the providers that can run: in a conversation, when you create an agent, on the routine canvas, and in the Slack and Telegram settings. If you can manage providers, a plus button in the picker opens the provider settings so you can add one.
In the full browser view, you now answer an agent's questions and approvals in the dock over the page. Before, the dock only told you to go back to the chat.
On the browser takeover card, Cancel is now on the left, and "I'm done" is the main button on the right. On a server, "Sign in on this computer" is next to "I'm done".

Fixed

An agent now uses its browser tabs for a web service that has no tool of its own, such as mail. Before, a routine run in its own session could not find a tool for the service and reported that it had no access, although a signed-in tab was open.
When an agent makes or changes a routine, it now writes an instruction that a run can follow without the chat: the service, the tool or browser tab, the account, the steps, and when to report. A run starts with only the newest chat messages, so an instruction such as "check the mail as before" did not tell it how.
Gemini Reconnect now opens the Google sign-in page, so you can sign in again or use a different Google account. Before, it checked the saved sign-in and nothing seemed to happen.
A long message in a card that asks if the agent can send it now scrolls, so you can read all of the message before you answer. Before, the card showed only the first lines.
A Codex chat now starts when ~/.codex/config.toml has an MCP server with the same name as an OpenBot MCP server of the other type, such as a local github server and the GitHub connection. Before, Codex showed "url is not supported for stdio" and did not load its configuration. The OpenBot server now goes to Codex as openbot_<name>, and your own server stays as you wrote it.
An OpenCode 2.1 turn that failed in OpenCode's local service now says so and tells you to try again or reconnect OpenCode. Before, the chat showed only "Internal error: Internal service failure".
A provider error that is not tied to a chat, such as a provider that did not start, now names the provider in its title, for example "OpenCode error". Before, the title was "Provider error", so you could not tell which provider failed.
An OpenCode turn that stops sending anything now ends after 5 minutes with an error that says so. Before, the chat could show "Working on it…" for an hour, and each message you sent again waited behind it. This also applies to Cursor, Cline, Grok and custom ACP agents. A running tool call or a question that waits for you does not count.
Stop now ends an OpenCode turn also when OpenCode does not stop it. Before, the turn could stay active after you pressed Stop.
An MCP server with a space or another special character in its name no longer stops Gemini and other ACP agents. Before, each task failed with invalid params. OpenBot now changes these characters to _ when it gives the server to the agent.
A reminder routine now sends its message in the chat when it runs. Before, the agent could decide that the reminder had nothing new to report, and the run completed with no message.
Show why channel routing failed, for example when the lead's model is unavailable, the lead did not answer in time, or the lead could not route the message. Before, the channel showed only "Routing failed. Choose a member or try again.", and no log recorded the cause.
A file link in a channel now opens the file in the preview panel, in the desktop app and the web client. Before, a click on a link to a shared file or to a file of an agent's workspace did nothing. A workspace link opens the file from the workspace of the agent that wrote the message. A quote of that file in a channel opens the same file again.
GPT-6 models in ChatGPT (Codex CLI) now call the tools of your MCP servers, such as Composio. Before, the server test found the tools, but the model did not see them and asked for another message. Existing conversations get the fix on their next message and keep their history.
Pi now shows as connected soon after you sign in with /login in Pi on the host computer, also over SSH on a server. Before, it stayed signed out until OpenBot restarted.
A hosted server that nobody uses for 10 days now starts for a short time and then goes back to sleep. This keeps its data: the hosting provider deletes the disk of a server that stays stopped for 14 days.
OpenBot no longer quits on macOS when an agent opens a page that redirects to another page while it loads, such as a sign-in or consent page.

Security

MCP servers that OpenBot starts with its own Bun now use Bun 1.4.4. It has security fixes for TLS, HTTP, WebSocket and fetch. OpenBot downloads it when an MCP server needs it.

Oct 22, 2026

Yaji 0.48.1

Fixed

Route channel messages again when the lead agent uses a free OpenCode model. Before, every message stopped with "Routing failed. Choose a member or try again.", because OpenCode refused the routing request. Agent profiles that a free OpenCode model writes work again too.

Oct 22, 2026

Yaji 0.48.0

New

In Server settings → Providers → Custom, each saved endpoint on this computer has an Edit button. You can change its name, base URL, API key, models and headers. Leave the API key and the header rows empty to keep the saved key and headers. Headers that you type replace all the saved headers. To use a different server address, type the saved key and headers again. The provider ID does not change. A joined server's endpoints do not have Edit yet.
Settings → Voice has a Remove button for each downloaded dictation model and for the speech model that reads replies aloud. Remove deletes the model files from this computer. A dictation model that you use downloads again at the next dictation. Replies are not read aloud until you download the speech model again.
Settings → General → Appearance has an Interface size control, from 67% to 200%. It changes the size of text and controls in OpenBot, and OpenBot keeps it after a restart. ⌘ + and ⌘ − (Ctrl on Windows and Linux) also change it, and ⌘ 0 sets it back to 100%. The View menu has Actual Size, Zoom In and Zoom Out.
You can choose a GIF as an agent picture, a profile picture or a server logo. An animated GIF stays animated. OpenBot stores it as an animated WebP.
The home page of openbot.run has the LaserFlow light beam from React Bits again. It falls onto the app preview and spreads along the top of the preview, in a soft violet haze. The light moves toward the pointer. The app preview has no shadow or glow around it.
The beam has a different colour, width, spread, haze and smoke on each visit. The smoke can be soft, or sharp with fine wisps. The colour is a blue, violet, pink or red, such as #3847ff, and is never yellow or green. With the reduced motion setting, the beam does not move.
The beam waits until the app preview has loaded, then falls in. The preview appears as the beam hits it. On a slow GPU, the beam uses a smaller canvas.
On a wide window, the beam falls through the Download button in the site header, and the button rim has the colour of the beam. The Community and App buttons in the site header and the "Available on" label in the hero are frosted glass. The App button is after the Download button.
The React Bits license does not let a public repository contain the beam, so it is in a private repository. A build from the public source code has no beam.
In a voice conversation, speak while the agent thinks or reads its reply to interrupt it. The reply stops, and what you say becomes your next message. The Stop button still works too.
Channels are back. The host offers channels to the web client again, and agents in a channel can assign work, transfer a task, post a result and read the channel history. Channels that the earlier release archived stay archived.
Agent import offers group chats as channels again.
In a channel, agents can show a plan card, an ask_ui block and an HTML page, and can capture the screen. Each screen capture in a channel asks first, also with Full access, and the request names the channel.
The desktop app and the web client notify you when an agent replies in a channel that you do not have in front of you. A click opens the channel.
A question, an approval or a browser step in a channel names the channel in its notification, and a click opens the channel.
The desktop app shows channels again on a server that offers them. Create one from the "+" menu in the sidebar.
Restore a deleted channel from its row menu in Deleted channels, on desktop and web. This also brings back the channels that the earlier release archived.
When OpenBot closes by itself, it now leaves a local record of the reason: a Quitting line in logs/main/main.log, an exit line and immediate process exit lines in logs/trace.ndjson, and a crash dump in the Crashpad folder when a process crashes. Nothing is uploaded. Do not attach a crash dump publicly: it can hold secrets.
Agents can list their open waits with openbot.list_waits, with each wait ID, note, open time and the routine that wakes it, and close a finished wait by ID. Before, an agent at the limit of 20 open waits could not find the IDs, so it could not open new waits.
openbot.list_agents shows the number of open waits of each agent. An agent can list and close the waits of another local agent with agentId.
When an agent is close to the wait limit, a new wait tells it to close finished waits. At the limit, the refusal lists the open waits.
Owners and admins of a joined server can now set an agent's backup models from the desktop app and the web client. Before, only the computer that runs the agent showed the setting, so a hosted server had no way to set it. Members do not see it. The server must also run this version.
You can install OpenBot with Nix on x64 or arm64 Linux and on Apple silicon: nix run github:nightly-labs/openbot. Nix installs OpenBot updates. A copy that Nix installed does not update itself; Settings shows that Nix installs its updates.
Gemini agents now show how full their context is, next to the send button. When the context is 80% full, OpenBot compacts it: the agent continues in a new session that starts from the recent conversation. The agent, its chat and its files stay.
The context popover has a "Compact now" action for Gemini and Codex agents. It compacts the context at once, below 80% too.
The Usage panel shows how much of your Gemini weekly limit is left, and when it resets. This needs a Google sign-in and Antigravity 1.4.1 or later: install the update that the Gemini provider page offers. Before, Gemini always showed "Not reported".
When a saved Gemini API key runs your Gemini agents, the Gemini provider page says so. A saved key always comes before your Google sign-in, so the work uses the key's project and not your Google AI plan, and the Usage panel shows no Gemini limit.
When your device's time zone changes, OpenBot asks if you want to move your routines to the new zone. It asks for this computer and for each joined server where you are an owner or admin, in the desktop app and the browser client. The routines keep their clock times, IDs and run history. Before, a routine always kept the zone it was made in, so after a trip it ran at the wrong local time.
A joined server now lets its owners and admins choose, from the mobile app, where an agent routine sends its answers: Slack, Discord or Telegram. The server must run this version.
You can ask an agent in chat to make a webhook routine, for example "start when my CI sends a failed build". The agent makes the routine and gives you its URL and the request format. A card in the chat shows the URL, and you make the signing secret there. The agent never gets the secret. The routine shows in the Routines settings with the other routines.
You can change which events start a webhook routine from the chat. Ask the agent, for example "only for the main branch", or change the event type and the filters on the card and press Save. The URL and the secret stay the same.

Improved

Check agent for a custom ACP agent now also shows the agent's models, so you can see what the agent offers before you save it. An agent that lists its models only after sign-in still passes the check.
A joined server can now show you agents on providers that your OpenBot does not know yet. The agent stays in your list and you can chat with it. Its provider and model choices are disabled and show "Update OpenBot to use" and the provider name.
New providers no longer need a new Team API version. Older OpenBot versions do not see agents on providers that they do not know.
Tasks opens in the main pane, as Routines does, and the server menu stays at the top. The agent list is hidden until you click the panel button in the Tasks header. A click on an agent opens it, and a click on Tasks goes back to the tasks.
Channel work holds only the agent that does it. Other agents and their tasks keep working, and an agent's own tasks run beside its channel turn. Only tasks that declare the same browser or folder run one after the other.
A channel in the desktop app now opens in the same chat view as an agent. Plan cards, visual replies, questions with choices and approvals show in the channel, and each message names its agent once for each run of messages.
A channel in the web client now opens in the same chat view as an agent, with plan cards, visual replies, questions with choices and approvals.
The Usage tab now shows tokens without cache reads in its total, chart, shares and daily data. Before, the total also counted cache reads. Each model request in a turn reads the whole cached conversation again, so cache reads made the total about ten times larger than the new tokens. Totals still show cache reads as Cached input, and the estimated cost does not change. For a remote server, the token chart shows one total line instead of one line for each provider.
Agents on ACP providers (such as OpenCode, Grok and custom agents) and on Muse now get their instructions once per session, not with each message. The instructions are 20,000 to 30,000 characters, so each message used thousands of extra input tokens. OpenBot sends the instructions again when they change, such as after a new memory, after the agent compacts its context, after a failed message and after a restart. An agent that does not report its context size gets them again every 10 messages.
A routine run in an agent's own chat now starts its own provider session, so it uses fewer tokens. Before, each run sent the whole chat context again. The run reads the newest chat messages, and its answer still shows in the chat. The agent reads the answer with its next message. A task, channel or messaging routine runs as before.
A phone notification of a question, an approval or a browser step in a channel names the agent and the channel, as on desktop. Before, it named only the agent. A phone app from before channels came back still gets the agent name.

Fixed

Images that an agent on a remote server attaches to a reply now show one by one. Before, a group of large images stayed as empty gray tiles until all of them arrived, and on a slow connection they sometimes did not show. A tile now shows a spinner while its image loads.
An agent now says that Marketplace has no app for a service, such as Whoop. Before, the agent said that it left a card to connect the app, but the chat showed no card.
Find and start a Codex CLI that npm, Volta, Bun or pnpm installed when you open OpenBot from Finder, the Dock, or a login item. Before, OpenBot reported that Codex was not found or could not start, because the CLI did not get your shell PATH. OpenBot now reads the PATH of the shell in $SHELL, such as bash or fish, not only zsh.
Say whether the downloaded copy or your own copy of Codex or Claude could not start, instead of always asking you to reinstall OpenBot.
On Windows, pick a runnable file when where finds a Claude, Codex or other CLI with more than one extension.
Usage analytics now count the tokens and the estimated cost of task turns. Before, a turn in a task stored no usage, for all providers. A Codex task that started before this version counts from its next request. A Grok task from before this version can count its earlier usage on its next turn.
Codex usage now includes cache writes at the cache-write price. Before, a cache write counted as uncached input.
A Codex request with 200,000 or more input tokens now gets an estimated cost. A request with more than 272,000 input tokens uses the long-context price.
A Grok update with less than 200,000 input tokens now gets an estimated cost when the session has more. Before, the session total had to be less than 200,000.
GPT-6.1 Sol, GPT-6 Sol and GPT-5.5 now have an estimated cost.
Delete an agent on Windows on the first try. Before, a delete often failed with "The agent data could not be removed completely" when the agent had saved a memory or had a profile change in the last 10 minutes, or when its Claude process took more than 1.5 seconds to stop.
Agent chats stay at the latest message while a turn runs. Before, the view jumped up and down, and the bottom was hard to reach.
A chat link to a folder in the shared directory now shows the folder in the file manager. Before, the link showed "Shared path is not a file." On a remote server, the error now says that the path is a folder (#2700).
An agent question that lets you pick more than one option now shows check boxes and a Submit answer button. Before, the first option that you clicked sent the answer. This works in the desktop app, in a joined desktop and in the web client of a server with this version.
An agent question that waits for your answer now shows again after the window reloads. Before, the question card went away and the agent waited with no way to answer.
A channel task moves to the agent's backup model at a usage limit, as the agent's chat does. The channel shows the model switch.
After a provider change, an agent's channel conversations get a handoff on the new provider.
A hidden lead no longer routes channel work, and hiding an agent gives its queued channel work back to the channel.
When an OpenCode model needs a sign-in that its provider does not have, the chat no longer shows only "Authentication required: provider authentication required". The error names the provider and tells you what to do: choose a model with the Free label, paste an OpenCode Go key in the OpenCode settings, or run opencode auth login in a terminal. The README says which OpenCode models need no sign-in.
When ChatGPT refuses requests because they come too fast, the agent now shows "Too many requests. Wait a moment, then try again." The raw Codex text "exceeded retry limit, last status: 429 Too Many Requests" no longer shows.
On Windows and Linux, Ctrl+W in the main window no longer closes OpenBot when no browser tab is open. With "Keep OpenBot running in the background" off, it quit the app with no question.
Answer the approval cards that already wait for an agent when you give it Full access, press Always allow, or an admin changes its permission from another computer. Before, those cards stayed open until you answered each one.
Run a Codex MCP tool without a card when the agent has Full access, also when the tool is set to ask every time in Codex. Computer Use still asks.
Stop showing thread not loaded in a Codex chat at each start. OpenBot now reads the history of an agent on a second Codex account through that account, and ignores a session that Codex has no saved history for.
Start a new Codex session when Codex has no saved history for the old one. Before, each message to that chat failed.
Agent and channel instructions longer than 2,000 characters no longer get cut silently. Before, a long paste dropped the rest of the text with no warning. Now the field keeps all the text, shows a character count near the limit, and shows an error when the text is too long. The instructions are not saved until you shorten them.
Claude usage now counts each turn once. Before, the first turn after OpenBot resumed a Claude session, such as after a restart or 10 minutes of idle time, counted all the earlier turns of the session again. The Usage view and its estimated cost showed too many tokens for Claude agents. Usage that was recorded before this version stays as it is. The first turn after a resume has no estimated cost.
When Google reports that a Gemini usage limit is reached, OpenBot no longer calls it a rate limit and no longer tells you to wait a few minutes. The message now says that the limit can be for one model, so the usage page of your plan can still show room, and it keeps the reset time from Google.
The menu that changes a routine's trigger now fits the space in the window and scrolls. Before, on a short window, such as one close to the Windows taskbar, the bottom of the menu was cut off and you could not reach Webhook.
Only the owner and admins can restore a deleted channel from another device, as they are the only people who can delete one. Before, the computer let a member restore it.

Oct 21, 2026

Yaji 0.47.1

New

Settings > Voice has a new local dictation model: Underdog Bark 3.0 Plus. It is fast and recognizes English only. It writes mostly lowercase text, with little punctuation. OpenBot downloads it (about 685 MB) only when you select it.
Roles in Server settings let an owner or admin choose the agents that members can use. Each role gives each agent one level: No access, Can view, Can chat or Can manage. A member gets the highest level of Everyone and of their roles. Owners and admins can use all agents. The role menu of each member on the Members page gives the roles and the server role.
Import your Hermes Agent profiles in Server settings › Import. Each profile becomes an agent with its persona, memories, skills and scheduled jobs. There are three ways, as for Grok Bot: - Read this computer: OpenBot reads the Hermes folder and shows every item, and why an item stays in Hermes, before anything changes. Skills with scripts and MCP servers import only when you allow them, and MCP servers are added turned off. When you import a profile again, its agent is updated, and no second agent is added. - Install with a command: type the /skills install command in a Hermes chat, send /openbot-export, and choose the .zip file that Hermes saves. - Set it up yourself: add the export skill file to Hermes by hand, then do the same steps.
Keys in .env and auth.json are never copied.
Server settings › Import now starts with where your agents are now: Hermes Agent, Grok Bot, or another OpenBot computer.
Select text in a file that the side panel shows, then quote it or ask the agent to change it. The quote names the file and, in a text file, the lines. Click the quote to open the file at the quoted passage.
Oh My Pi (omp acp) is now in the list of known agents when you add a custom agent, and the agent scan finds an installed omp command.
Ask an agent to add an MCP server, for example "Add this MCP server: https://example.com/mcp". The agent connects to it, tells you its tools or the error, and you approve each change unless the agent has Full access. When the server needs a sign-in, a Sign in card shows in the chat; press Continue after you sign in, and the agent tests the server again. Keys stay in Server settings, MCP: the agent never asks for them in the chat. Agents can also turn MCP servers on or off, test them and remove them when you ask.

Improved

Settings > Providers is now a short list in two groups, Connected and Other providers, with a search field. Each row shows one line, and a badge only when the provider needs you.
Each provider has its own page with its accounts, sign-in, API key, version, restart, diagnostics and on/off switch. Open it from the list; "Providers ›" in the title goes back.
An account that is low on usage shows a badge with the amount left. Click it to see all of its limits.
The default provider for new agents is now one menu at the top of the list.
Custom endpoints have their own page, where you add and remove them.
A member can no longer change an agent's settings when a role gives that member less than Can manage. On a server with no roles, members can use all agents, as before.
The Quote button in the selection bar shows its shortcut in its tooltip.

Fixed

The Tasks page keeps the agent that you chose last on each server. Before, it went back to the first agent each time you opened the page.
The agent menu on the Tasks page scrolls when a server has many agents. Before, agents below the bottom of the window could not be chosen.
Icons in buttons with text are now the size of the text, not larger.
OpenBot now offers new Pi releases. Before, the update check could not read Pi's release tags, so Pi stayed on the version that came with OpenBot.
In Discord, the agent now gets the images and files of earlier messages. A new conversation also gets the channel's messages from the 15 minutes before it. Before, "@OpenBot make an issue" after an image that another person posted without a mention did not see that image, and the agent could act on an older request.
Keep each agent's provider session when OpenBot stops through a signal, as a terminal, systemd or Docker sends it. The Computer Use driver stopped from the same signal and made every agent start a new provider session at the next launch, which lost the provider's own context.
The Composio listing now connects with a Composio Connect link (https://connect.composio.dev/mcp) and its API key. Before, the key went in a header that Composio Connect does not read, and the connection failed with "This server asks for a sign-in".
A server that refuses a pasted key now says to check the key. Before, it said to sign in.

Oct 21, 2026

Yaji 0.47.0

New

The model picker has a Fast switch for Codex models that have a fast service tier. When it is on, the agent asks Codex for the fast tier and the picker shows "Fast" next to the model.
Attach ZIP archives in chat. The agent gets the archive unchanged and can extract it. An agent can also send a ZIP archive back, and you can download it from the chat. A remote host must run this version of OpenBot to accept ZIP archives.
Add Inception as a provider, with the Mercury 3, 2.5 and 2 models. Select Connect on the Inception row and save an Inception API key. Inception runs in the Pi CLI that OpenBot downloads, so download Pi first. OpenBot keeps the key encrypted on this computer and does not change Pi's own settings. Inception agents stay on this computer and are not shown to team members.
The desktop app and the web app show a notification when an agent needs you to finish a step in the browser, such as a sign-in or a code. Before, only the phone and the MacBook notch showed it, and the phone stayed quiet while you used the desktop app.
A message or a routine run that cannot start now shows a notification on desktop, web and phone, as a turn that stopped with an error does.
Read the SKILL.md of a skill on a joined server and in the web client. The text comes from the host. Before, a local skill or a folder skill showed only its description, and a skill that was changed on the host showed the marketplace text.
On a joined server, a skill marker in the chat opens that skill in the agent settings.

Improved

OpenBot now installs OpenCode 2 (2.1.1) as its own OpenCode CLI. Your threads keep their OpenCode sessions.
When you mute this computer's server, or set its notifications to Nothing or Only when it needs me, your phone now gets the events that the desktop does not show. Before, these events reached no device while you used the computer.
Mute and the notification level of a server now also stop the agent sounds on desktop, as they do on web.
You can change the model and effort of an agent while it works. The running turn keeps its model, and the new model starts with the next turn. An "After this turn" badge beside the model picker shows that a change waits. Before, the picker was locked until the work finished. A change of provider or account still waits for the work to finish.
Server settings: Publish this server is now in Members, above the invitations. General keeps the server name, logo, and preferences. The server address row is removed: people join with an invitation, not an address.
You can publish a server before you name it. OpenBot gives it a name from your account, such as "Ann’s server". You can change the name in General.
When product analytics is on, a completed agent turn now also sends the token counts that the provider reported for that turn: uncached input, cached input, cache write, output, and total. It also sends these counts for each provider and model that the turn used, such as a subagent's model. When the provider reported no usage, it sends that the usage is unknown, not 0. No message content is sent. You can turn product analytics off in Settings → General → Permissions → Share product analytics.
Product analytics now also counts the hosted server plans, plan and member limits, license steps, contact sales links, provider usage limits, setup steps, connectors, agent templates, phone connection, channels, server export and import, help links and the Usage view. It sends no names, message text, IDs, links, prices or email addresses.
Licenses now appear in the billing analytics of the account service, with the number of users.
The web app at openbot.run/app now sends the same product analytics as the desktop app. Its setting is now Share product analytics in account settings.
When you turn off Share product analytics, the app sends one last event that records the change. Then it stops all analytics.

Fixed

Center the hero of the openbot.run download page, as on the other site pages. The title is now the same size as the other page titles, so it shows on one line on a wide screen.
A remote session stays connected when Signal restarts and the account service is slow to answer. Before, the session closed and the device showed it as offline.
After a Signal restart, OpenBot computers and devices no longer all connect again at the same time.
OpenCode setup no longer fails with "could not activate the managed CLI" on a computer where OpenCode 2 is already installed. When OpenBot uses an older OpenCode CLI that cannot open the database of OpenCode 2, setup now tells you to update OpenCode to version 2 or later.
With OpenCode 2, an OpenCode thread opens again after a restart, also while OpenCode is still starting.
With OpenCode 2, the model picker shows the OpenCode, OpenCode Go and custom endpoint groups by their names.
Remote chats no longer show a permanent chat error when a background queue refresh times out. The last queue stays visible until the next update. Failed queue actions still show an error.
The Voice select in Settings → Voice → Spoken replies now shows the voice you choose. Before, it kept the name of the previous voice until Settings opened again.
Sign in to an MCP server now reports an error when the sign-in information of the server cannot be read. Before, the press ended with no page and no words, and the row kept the result of the earlier test.
A reply quote on desktop shows the Markdown of the original message: headings, bold text, code, lists, links and code blocks. Before, it showed the raw marks, such as ## and backticks, on one line. The quote now shows up to four lines, in the composer and in the sent message.
On a phone, the web side panel shows its tabs in a row under the chat header. Before, the tab row covered the agent name, the model picker and the side panel button, and with the pane list open the side panel button did not respond.
A Codex chat no longer shows the error "Falling back from WebSockets to HTTPS transport" when Codex changes its connection. The reply continues. If the reply then fails, the chat still shows the error.
Center the agent avatars in the connectors list of Server settings and on the Slack, Discord and Telegram connector pages. Before, each avatar stood about 3 pixels lower than the row text and the arrow.
OpenBot now keeps your sign-in, API keys and other secrets on Linux desktops such as Hyprland, Sway, i3 and LXQt when a Secret Service (such as GNOME Keyring) runs. Before, it asked you to sign in on every start, and it could not save a pasted API key.
Keep your Codex and ChatGPT thread lists clean. When OpenBot starts a new Codex session for an agent, for example after you change its tools or reset its context, it now archives the session that the new one replaces. Before, each replaced session stayed in your Codex history as a row that starts with "Continue this OpenBot conversation." The agent's conversation in OpenBot does not change.
Pi agents keep their OpenBot tools, browser included, in new conversations. Pi also shows the OpenBot tools with their real names, so a model can find them.
A routine run in the chat now animates once. Before, its row and its steps, such as "Invoked" and "Started", could animate again when the run changed state, when two runs overlapped, or after a reload.
When an OpenCode model cannot read a file or image in the chat, OpenBot now tells you to choose a model that can read files, in a yellow notice. Before, a red error said that the provider failed and that you must try again later, but each new message failed again until you changed the model.
A task that needs your approval no longer looks stuck in "Making the requested changes…". The task panel on desktop and on the web shows the approval with Allow and Deny, and the task says "Waiting for your approval". An approval of a task no longer hides an approval of the agent's chat.
On Discord, OpenBot reads the images and files of the message that you reply to. Before, it read only the text of that message, and a message with only an image gave no context.
A notification for a question, an approval or a browser step now goes away when the request is answered or cancelled, also from another device. Before, it stayed in the Notification Center.
When the ChatGPT sign-in of a server had expired, a sign-in with a code from the web client or the phone stopped a few seconds after the code appeared. The code now stays valid until you enter it.
Failure reports from the desktop app and the local host were dropped before they were sent. They now reach the analytics service.
Check saved secondary ChatGPT and Claude accounts when their provider starts, so accounts can reconnect after an app reload or CLI update without removal and a new sign-in.

Oct 20, 2026

Yaji 0.46.0

New

OpenBot can read replies aloud on this computer. Turn on Read replies aloud in Settings › Voice › Spoken replies. When you dictate a message, its reply is read aloud in the language of the reply, in 31 languages. Dictate again, or send a new message, to stop it. Choose one of ten voices and press Listen to hear a sample. The Supertonic 3 voice model downloads once, about 145 MB, and no text or audio leaves the computer.
A bar above the composer shows the reply that is read aloud, with a Stop button.
Each agent message has a Read aloud button. Press it again to stop reading.
Talk to an agent in the Voice pane of the side panel. Open it from the side panel or from the "+" menu of the composer. Speak, and stop talking to send the message. The agent reads its reply aloud, and then the pane listens again. The first time, the pane downloads the voice models. A waveform under the agent moves with your voice, so you can see that it hears you, and with the agent's voice while it speaks. Mute, Interrupt and End control the conversation.
A message that you speak in the Voice pane has a voice player in its bubble, with the waveform and length of your recording. Play, pause, or click the waveform to go to a different part. The agent gets only the text, and the recording stays on this computer. The reply to a spoken message has a player too: it reads the reply aloud with a timer, and you can pause and resume it. After one full reading, the reply keeps its audio until you close OpenBot: its player shows the waveform and length, plays again at once, and you can click the waveform to go to a different part.
The Voice pane works with an agent on a remote server. The spoken reply runs on this computer, speech recognition uses the voice model you chose in Settings › Voice, and only the text goes to the server. A message that you speak to a remote agent has no voice player.
When the server of the agent disconnects, the Voice pane stops listening and tells you. It listens again when the server connects. When a spoken message cannot be sent, the pane tells you and listens again; the message stays in the chat with Retry, Edit and Dismiss.
Voice recordings use WebM/Opus, about 4 KB a second. They show in the Storage tab of Server Settings, where you can delete them, and they are part of a data export.
The production desktop app sends OpenPanel events for the Voice pane, spoken replies, the voice model download and the spoken reply settings: counts, times, lengths and results only. It never sends what you said, the reply, or audio.
Invite users to the OpenBot community from a card at the bottom of the sidebar, in the desktop app and the web client. "Join the Discord" opens the Discord server, and the X button opens the OpenBot account on X. The card stays closed after you dismiss it. On desktop, it shows after you dismiss the mobile app card.

Improved

Set "Memories per agent" only in Server settings → General → Agents, for this computer and for joined servers. App Settings no longer shows it. Before, the App Settings value applied only to agents on your computer, so a server's agents could stop at 64 memories while App Settings showed a higher limit. Global search also finds "Memories per agent".
The top of the openbot.run home page no longer shows a grid behind the title.
In the web app, an account with no server now sees Create a hosted server as the first action. Download OpenBot and Join with invitation are still there.
Show the License section in Billing also when licenses are not for sale. A purchase then shows a message with the address to contact: licenses@openbot.run.

Fixed

Keep the saved account sign-in when OpenBot cannot read or save it at one start. Before, OpenBot deleted it, and you had to sign in with an email code again after the next restart or update. OpenBot now writes the reason to logs/account/account.log in its data folder.
Keep a question from a ChatGPT agent open until you answer it. Before, with GPT-6 models the question often expired after about 20 seconds.
On Windows and Linux, "Keep OpenBot running in the background" now works. When it is on, closing the window hides OpenBot, and an icon in the system tray opens it again or quits it. Before, the window closed OpenBot, and the switch was off again after a restart.
"Launch OpenBot at login" now opens OpenBot when you sign in to your computer. Before, the switch did nothing and was on again after a restart.
OpenCode agents no longer fail on each message with "OpenCode service failure" when OpenCode cannot read the agent's session, for example a session that another OpenCode version wrote. The agent now continues in a new session with its conversation.

Oct 20, 2026

Yaji 0.45.0

New

One OpenBot server can connect more than one Slack workspace or Discord server. Settings › Connectors shows "Add workspace" or "Add Discord server" after the first one is set up. One Slack Orchestrator answers every Slack workspace, and one Discord Orchestrator answers every Discord server, as one Telegram Orchestrator already answers every Telegram chat.
You can choose which Discord roles can use OpenBot in each Discord server. OpenBot ignores a mention from a member who has none of the roles. With no role chosen, everyone can mention OpenBot, as before.
The context meter in the composer has a New chat button. The earlier messages stay visible, and each new message sends less context.
An agent can start a new chat for itself with the start_new_chat tool, for example from a routine. It saves notes first. Turn on "Agent starts new chats" in the agent's settings to allow it. It is off by default.
Ask side questions in the web client and on joined servers. Before, side questions worked only on the computer that runs the agent. The server must have this version of OpenBot.
An agent can show its plan before a long task. Turn on "Show the plan before long tasks" in the agent's settings. The agent then posts a short card with its plan and Start, Cancel, and sometimes a smaller option. The agent starts after 10 s if nobody answers. It does not show the card for quick answers. The setting is off by default.
The chat shows the subagents that a Claude or Codex agent starts. One card for each turn shows how many subagents finished, what each one does now, and how long it took. Click a subagent to see its task and all of its steps in the side panel.
When you make a routine on this computer or on your own remote computer, OpenBot offers an OpenBot server, which runs routines 24/7. The offer shows after an agent's turn ends, at most once in 30 days. The routine list and the offline notice of a remote computer also show the offer. "Not now" hides it for longer each time.
Settings → Billing has a License section for work at a company. Buy a license for 5 to 25 users, in EUR, USD or PLN. See who uses it, change the number of users, update the payment method, or cancel it. A larger team can contact us.
A license limits the people on the owner's servers. When it is full, a new person cannot join. The server owner sees the extra price at a new invite and can add 1 user.
A member of a server whose owner has a license sees that the license covers them.

Fixed

On Windows, OpenBot now finds an OpenCode that you installed with the OpenCode installer or WinGet. Before, it said that OpenCode was not downloaded.
On Windows, the Download button of a provider works again after a shutdown or sign-out that was cancelled. Before, each click failed and showed nothing.
In Settings, a provider download that fails now shows why. Before, the Download button did nothing.
The Cline model picker shows Cline's free models, such as MiniMax M3.1 Flash, MiMo V2.6 Flash and Muse Spark 1.3 Contributor. Before, only the Cline CLI showed them.
A Cline agent runs on the model that you select. Before, the agent stayed on Cline's default model.
Steer on a remote server now shows why it did not steer when the provider, such as OpenCode or Grok, cannot steer a running turn. Before, it showed only "Request failed". The message still stays in the queue.
When you mention OpenBot in a Discord reply to someone else's message, OpenBot now reads the message you replied to. Before, OpenBot said that it could not see what "this" was.
The delete server dialog keeps its height and position when you change between "At period end" and "Immediately". Before, it moved because the description and details have different lengths.
A Claude agent now makes a routine in OpenBot when you ask for one. Before, it sometimes used Claude's own schedule skill and made a claude.ai cloud routine that OpenBot did not show. Claude's own scheduling tools are now off in OpenBot.
A chat with unread messages now opens at the "NEW" divider, with the last read message above it. The divider stays while you read and goes when you leave the chat. Before, the chat opened at the newest message and the divider went away at once, because the open chat marked the messages read.
An OpenCode agent now uses the OpenCode Go key that you paste into OpenBot. Before, an old key that opencode auth login saved took its place, so a turn failed with "This API key has expired" again after you replaced the key.
An expired OpenCode key now shows "Your OpenCode API key expired" with an Update key button that opens the provider settings. Before, the chat showed "Internal error".
A Workspace-only agent now uses a new provider key on its next turn. Before, it could keep the old key until it was idle for 10 minutes.
A cloud server that started again soon after it stopped sometimes stayed at "Starting Cloud server" with signal/host_unavailable. The machine continued from memory, and OpenBot did not start on it. The account server now starts OpenBot after such a start.
The update banner now shows the full button label in every language. Before, a long label such as "Uruchom ponownie" (Restart) in Polish went past the button and the banner cut it off. In Polish, the button now says "Restart". In Turkish, the "managed by host" button now says "Ana makine yönetiyor".
Renew in Settings → Billing works again for a server plan that you cancelled. Stripe refused the request before.

Oct 19, 2026

Yaji 0.44.3

Improved

The rail on the right side of a long chat now shows your messages, not the days. Point at the rail and select a message to go to it and to the replies under it. The rail is there when a chat has two or more of your messages.
The rail of a long chat stays in view when you do not scroll. When you point at it, the chat behind the titles is blurred, so the titles are easier to read.

Fixed

The results of the settings search now show their text on the left. Before, the section name and the row title of each result were in the center of the row.
The rail of a long chat no longer shows days out of order or the same day two times.
The rail of a long chat stays smooth while you scroll, and its titles no longer come in slowly in a long list.
"Side question" in the composer "+" menu now opens the side question pane. Before, a click with the mouse closed the menu and did nothing.
An agent template that contains the OpenBot skills openbot-data, openbot-site-hosting or openbot-skill-creator installs again. Before, the install stopped with an error, because the new agent already has these skills. A new publish no longer adds these skills to the template.

Oct 19, 2026

Yaji 0.44.2

Fixed

Show the whole tab menu in the browser view. Before, a right-click on a tab showed only the first rows of the menu, and the page covered the other rows. Now the page is hidden while a menu of the browser view or the address suggestions are open, and it comes back when the menu closes.
The page of a shared agent, and the dialogs that add or publish an agent in the desktop app and the web app, show headings, lists and links in the agent's instructions. Before, they showed the Markdown marks, such as ## and -, as text, and the page centered the instructions.
You can now scroll up in a chat while the agent works on a reply. Before, a short scroll up with a trackpad or a mouse wheel moved the chat back to the newest text each time the reply grew, so the chat jumped and you could not read the earlier messages. The chat follows the reply again when you scroll back to the bottom or send a message.
Let an agent open pull requests and post comments with gh when GitHub is not connected. Before, the agent refused and asked for GitHub MCP tools that it did not have.

Oct 19, 2026

Yaji 0.44.1

New

Set "Memories per agent" for a joined server in Server settings → General → Agents. Before, the setting in App Settings applied only to agents on your computer, and agents on a self-hosted server kept the limit of 64. You must be an owner or admin, and the server must run this version.

Fixed

Start OpenBot when a power loss left the joined server list or the team server settings file empty. Before, the start stopped with the error "Unexpected token '". The damaged file stays beside the new one with a .corrupt-<time> name. OpenBot also flushes each settings file to the disk before it replaces the old one, so a power loss cannot empty the file.
Name the damaged file when a settings file stops the start. Before, the error quoted the start of the file and did not say which file it was.
Open the file panel for a local file link over 100 MB. Before, a click on the link showed the error "The file exceeds the 100 MB limit." and no panel opened, so you could not open the file, show it in the folder or save it. Now the panel shows that no preview is available and gives these actions. OpenBot does not read the file to show the panel.
Let Claude agents run tools from Homebrew, ~/.local/bin and other folders of your shell PATH when you open OpenBot from Finder, the Dock, or a login item. Before, the Bash tool and hooks reported command not found. An env.PATH in your Claude settings still wins.
Keep the server and its members connected when an owner or admin changes the role of a member. Before, each change to Admin or Member disconnected the host and every connected member, and the server showed as offline until it connected again. Now only the member whose role changed connects again, with the new role.
Show the full day tooltip on the status page. Before, the card cut off the tooltip of the first and last days.
Run a message again to its end after the provider refused its first turn. Before, the second run of the message showed as interrupted while a Claude, ACP or Pi agent still worked on it. A Claude agent then refused each new message with "The Claude thread already has an active turn." until OpenBot restarted.
Move the window by dragging the top row of the Tasks page. Before, only a thin strip beside the "Tasks" title moved the window.
Open Settings again. In 0.44.0, Settings showed an empty page and the app stopped responding until it restarted.
Update a provider CLI while the provider works on a turn. Before, Update and Retry showed "Update failed" with "The Claude CLI is working on a turn. Wait for it to finish, then update.", and you had to press Retry again later. Now the provider row shows "Waiting" and the update starts by itself when the current turn ends. New turns of that provider wait for the update. Cancel stops the wait.
Let a Claude agent answer again after you switch it to OpenCode and back. Before, each message failed with "--resume requires a valid session ID", also after a restart. An agent that has this fault now starts a new Claude session and keeps its conversation.
A provider account that is signing in now says to finish the sign-in in the browser, and its menu has Cancel sign-in. Before, the account showed "Connecting" with only Rename and Remove until the sign-in timed out.
A new account sign-in on a joined server's host now stops the earlier one. Before, the earlier account could stay "Connecting" for up to 10 minutes, and Cancel could not stop it.
Show the answer of a scheduled, script or webhook routine run one time in the chat of a Claude agent. Before, after OpenBot restarted, the chat showed the same answer two times in the run.

Oct 18, 2026

Yaji 0.44.0

New

Agents can ask with interactive blocks: a confirmation with buttons, reply chips, a list to choose from, or a short form. The agent waits for your answer, as it does for a question. Every provider can use them, Claude too. The blocks show as cards in the desktop app and the web app, and you can skip one as you skip a question. Only the server owner or an admin can press a dangerous button, or answer a block that has one in words. A form can ask for a password or a key: the agent gets it, but OpenBot does not keep it on the card or in your answer. The agent can still repeat it in its reply or its thinking. If you cannot hold a button, press it and confirm in a second step. An app or a mobile app that does not show the blocks yet shows them as an ordinary question that you can answer.
Settings > Profile shows the activity of this computer in the past year: a calendar of active days, the tokens, the messages, the peak day, the streaks, the providers, the models and the teammates that you use most.
Share in Settings > Profile makes an image of your last six months. You can copy the image or save it as a file. The image stays on your computer until you send it.
Add a Tasks page to the page rail, on desktop and on the web. Write a task, choose the agent, and send it; you can undo the send for 6 seconds. The page groups the tasks of all agents by what they need: Needs you, Working, Waiting and Closed. Click a task to open its conversation beside the list. On desktop, you can also dictate a task. A joined server shows the page when it supports tasks.
Take over a browser tab from its agent at any time, and give it back with "I'm done". The agent waits while you control the page, then continues from a fresh look at the page.
Right-click a browser tab to open a new tab to its right, duplicate it, move it, or close the tabs to its left, to its right, or all other tabs. Reopen a closed tab at its old place.
The address bar suggests open tabs, pages you visited in this session, and a search as you type, and it completes a visited address.
Stop a page that loads from the address bar. Back and Forward are off when the tab has no page to go to.
In the full browser view, the keys of other browsers work: new tab, close, reopen, select a tab by number, next and previous tab, address bar, reload, back and forward.

Improved

Signing out of an MCP server, or removing it, now asks the server's authorization service to revoke the tokens.
Agents no longer get the OAuth token of an MCP server that you signed in to. They reach the server through a local address in OpenBot, which adds the token to each call.
Save on a new HTTP MCP server now tests it first. When the server asks for a sign-in, OpenBot signs in and then saves. When the server refuses the typed credentials, OpenBot does not save and goes to Headers.
A web address pasted or typed into "Command to launch" now moves to Server URL and selects Streamable HTTP.
Profile is now the second tab in Settings, after General.
You change your photo and your display name in one Edit profile dialog, in the desktop app and in the web client.
Reply on a message in a chat or a channel now puts the caret at the end of the message box, so you can type at once. Before, you had to click the message box first. Text that you selected in the message box stays selected.
Agents with a long message history start queued work faster.
The web app uses less memory when you open many long conversations.
The full browser view shows every tab in its top row, with the address bar above the page. In a narrow window, one button with the tab count opens the list of tabs.
The composer, the queue and the agent status float over the page in the full browser view, so you can talk to the agent of the tab while you look at the page. The agent's reply shows over the page and closes by itself.
Select a tab of a different agent in the full browser view to open that agent's chat. The page stays open, and the floating composer sends to the agent of the tab.
The card where an agent asks for a code or a password has a new design.

Fixed

Open the sign-in page when you click Sign in for an MCP server that shows its tools without a sign-in, such as the Google Sheets and Google Chat MCP servers. Before, the click did nothing, and the server stayed signed out.
Connected devices in Mobile Connect now shows only the phones that can open this computer. A phone that you connected to a joined server no longer shows on a computer that it cannot open.
Keep messages with unknown code language labels visible when a label is constructor or __proto__.
Wrap a long agent reply inside its bubble when it names a file by a long path. Before, the text ran past the right edge of the bubble and was cut off, and the chat column could grow wider than the window.
Remove temporary instance transfer ZIP files left after a restart when the next upload or export starts.
Complete marketplace skill installation and refresh the agent runtime when the install receipt request fails after local files and the skill lock are saved.
Prevent service restarts from interrupting a teammate file download while its input stream is still being written to disk.
Routines now use the earliest next scheduled time when clocks move back. Some schedules skipped a run in the first occurrence of a repeated hour or missed a run when the clock moved back several hours.
Reject malformed custom cron list and step syntax when saving a new or changed routine schedule. Keep saved schedules readable and editable.
Keep all attached files when the file system treats different filenames as equal. Previously, one attachment could overwrite another and become unavailable.
An MCP server that refuses a saved sign-in now shows "Sign-in needed" with a Sign in button. Before, its row looked signed in until a test failed.
A background token renewal no longer registers a new OAuth client with the MCP server.
Agents no longer get an MCP token that the server has already refused.
An MCP token that expires during a long agent session is now renewed. Before, the agent kept the old token until its session restarted.
After you sign out of an MCP server, or the server ends the sign-in, an agent that calls one of its tools now gets a message that tells it to sign in again in OpenBot.
After you cancel a checkout from the desktop app, the billing page now says that the checkout was cancelled and that nothing was charged. Before, it said "Billing updated".
The form to report a hosted site now has the site style and shows a clear result or error. Before, it had no style and showed raw data after you sent it. You can now type the site address when the form opens without one.
"New URL" and "Turn off" for the routine calendar feed now ask before they stop the current URL. The panel also shows a loading state, a Retry button after a failed load, and a note when the agents cannot load.
The routine calendar no longer shows channel routines that cannot open. Before, a click on one could leave a pane that loaded without end.
The first-agent setup screen and the side-question panel now show in your language.
When a side question gets no answer, the panel now says that it failed and has a Retry button.
Reject custom cron numbers that are not plain decimal digits, such as 0x5, 1e1, +5 or 5.0, when saving a new or changed routine schedule. Keep saved schedules readable and editable.
Complete the installation of a marketplace agent when the install receipt request fails after the agent is saved.
When a setting does not save, a message now tells you why the switch went back. When settings do not load, the settings page says so and has a Retry button. Before, it showed default values.
When OpenBot cannot read your saved setup at start, it now shows Retry. Before, it opened first-run setup.
1Password, Bitwarden, Slack, Discord and Telegram now show "Unknown" and Retry when their status does not load. Before, they showed "Not set up". A failed "Open app" for 1Password now shows an error.
Slack and Discord now have Pause, as Telegram does.
Disconnect for an account session now asks first.
You can undo the delete of one memory.
Each routine in the list now has Pause or Resume. "Delete now" shows that it is in progress.
Failed actions in the browser picture-in-picture controls, the What's New link and the routine webhook status now show an error.
The join, Slack, Discord, billing, site report and error pages on the website now have the same header, card and footer. The page-not-found page has its own title.
The Slack and Discord connect pages now say what went wrong for each error and offer "Try again" and help when OpenBot does not open.
The join page now says "server", explains how invitation links work, and has link previews.
Download links on the website now open the download page. The footer now links to Teams and Contact.
A shared agent that is no longer shared now shows a clear page, not a generic error.
Hosted sites now show a branded error page in the browser, and their report link uses the correct account service address.
The web app can now create a hosted server when you have no server to connect to.
Stop the provider processes when OpenBot quits while it changes a provider key, updates a provider CLI or reconnects a provider. Before, the new process, or the process that it replaced, could continue to run after OpenBot stopped.
The app and the web app now use "server" for a computer that you connect to or join. Before, they also said "host" or "computer".
Error labels in a conversation now say what failed, for example "Could not send your answer". Agent errors now tell you what to do next.
Errors in the agent registry now name the action that failed: search, install, update, remove or cancel. Before, all of them said "The registry operation failed".
The join, Slack, Discord, billing, site report and error pages on the website now show in the language of your browser.
More text in the web app, the sidebar, the marketplace and the Cobrain listing now shows in your language.
Settings search now finds settings on every tab. A result opens the tab and marks the setting.
Icon button tooltips now use the app style and close with Escape. Before, some used the plain system tooltip.
A button that is busy now keeps keyboard focus.
All dialogs now use the same dimmed background.
The web client and a joined desktop can now read the queue of an agent with a long history on a server. Before, after some weeks of use, the queue answer was too large for the connection, and the chat showed a "response too large" error (#2368).
Show the correct reason when a direct message cannot go to its recipient, for example when the member left the team or is disabled. Before, the server answered "Request failed." and wrote an error to its log.
A message no longer stays "running" after its turn ended with an error. Before, a failure at the end of a turn or during the recovery after a restart could keep the message running, and the agent's next messages waited in the queue. Stop did not clear it. Now the message ends as interrupted or failed, the next message starts, and Stop clears a message that no provider runs.
PRIVACY.md now lists the hosted site files, marketplace skill bundles and site reports that the account service stores.
Opening a server invitation link again no longer disconnects every member of the server. A member keeps their role, and a removed member cannot join again with an old link.
When a provider process stops, agents on other processes keep their browser controls, and a compaction of the stopped process shows as failed instead of staying hidden.
Signing out, or ending a session from a phone or the web, now also closes the live browser view of that session.
The Live Activity updates that OpenBot sends to your iPhone show a link that an agent reply writes in angle brackets, such as <https://example.com>. Before, the update removed the link.

Security

An MCP sign-in now stops an OAuth answer that is larger than 1 MiB or slower than 20 seconds.
A public MCP server can no longer send its sign-in to an OAuth endpoint on a private network.
A member that the owner or an admin removed could rejoin a server with an invitation or a team link from before the removal. Now that member needs a new invitation or a reactivation. A member who left by their own choice can still rejoin with an earlier invitation. A member who left before 18 October 2026 needs a new invitation, because OpenBot did not record before then who ended a membership.
An invitation from an admin stayed valid after the admin left, was removed or was demoted. Now it closes, and it no longer counts against the invitation limits.
If the public ticket key set had a private key by mistake, /.well-known/jwks.json published the private key. Now it publishes only the public members.
A checkout return address that starts with // or /\ could send the browser to another site. The account server never made such an address, but now it refuses one.
A hosted site address such as /constructor made the site router read storage for a file that does not exist. Now the router finds only the files of the site.
A request with no Cloudflare client address got its rate limit key from X-Forwarded-For, which the client sets. Now only the Cloudflare address sets the key.
An IPv6 client got a new per-address rate limit for each address in its /64. Now all addresses of one /64 share a limit.
Signal read the full body of a Slack request or an account event before it checked the size. Now it stops reading at the limit and answers 413.

Oct 17, 2026

Yaji 0.43.0

Action needed after you upgrade

Sign in to Vercel and other OAuth MCP servers again after you upgrade.

New

Add "Join the Discord" to the Help menu of the page rail, on desktop and in the web app. It opens the OpenBot Discord community in the browser.
You can hide an agent that you do not use now, and keep it. Select "Hide agent" in the agent menu. The agent moves to the "Hidden" section at the end of the agent list. Its thread, files, memories, settings, section and pin stay. A hidden agent does not run its routines and does not get new messages. Messages that are already in its queue wait. Select "Restore" to put the agent back in its section. Routine runs that it missed while it was hidden do not run. The hidden state is the same on each device that shows the server. On a joined server, only an owner or an admin can hide or restore an agent.
Add a Month view to the Schedule page. It shows the runs of each day of a month in whole weeks, with three runs on each day and a count of the other runs. Select a day or the count to open that day.
Add a "Provider runtimes" item to Clean up on the Storage page of this computer. It removes old provider CLI versions at once, and it does not wait for the automatic cleanup.
Show what fills a Claude agent's context in the context popover of the composer: the system prompt, memory files, tools, skills and subagents, tool calls and results, messages, and free space. The counts are Claude's own estimates and need no extra model request. Other providers still show only the total. (#1635)
On the website, when the agent network fills the frame, click an agent to grow a new network from it.
On the website, click the OpenBot logo in the agent network to send a signal to the agents. Each click shows one of three signals at random.
On the website, the agent network has easter eggs. Type "hello", "party", "sudo" or the Konami code, click the hub quickly many times, or leave the page alone until the agents fall asleep. Each reaction chooses its words and signals at random. Reduced motion turns the easter eggs off.
On the website, agents in the network chat about their work when you click the hub, and when a network that you regrow is complete.
Set the size of the virtual display that computer use reads on a self-hosted server, with --display-size at the install or OPENBOT_DISPLAY_SIZE in /opt/OpenBot/hosted/openbot.env, and in Docker with the OPENBOT_DISPLAY_SIZE variable. The default stays 1280x800.
Ask the agent a side question from the composer "+" menu. The agent reads the chat and answers in the side panel. The question and the answer do not go into the chat, and closing the tab deletes them. Side questions work only on the computer that runs the agent.
Add a public status page at status.openbot.run. It shows the state of sign-in, connections, hosted servers, desktop update downloads and the Marketplace, 90 days of history, incidents and planned maintenance. It has an RSS feed, and the site footer links to it.
Click the quote in a reply to go to the original message. The chat scrolls to it, loads older history first when necessary, and highlights the message for a moment. "Back to the reply" takes you back. When the original was deleted, the reply says "Original message not available".
Quote selected text from a chat message in a new message. Select the text and choose Quote, or press ⌘⇧. on macOS or Ctrl+Shift+. on Windows and Linux. The quote goes into the composer as a chip at the caret, and one message can have more than one quote. Click a quote to go to the quoted text, which shows highlighted. The agent gets each quote as a Markdown block quote.
The status page sends signed webhooks for new incidents, updates, maintenance and service check changes to the tools of the OpenBot team, with retries for 11 hours. Slack and Discord formats are available. The feed stays the public way to follow the page. (#2335)
openbot-cli runs a task on an agent from a terminal or a script and can wait for the answer. It also changes an agent's model, reasoning effort or provider account, and starts a new chat with an agent. It uses the Local scripts API, so only agents with Local scripts on can be used. When OpenBot restarts during a wait, the command continues the wait. See docs/automation.md. Thanks to @NatanPimentel.
The Local scripts API has three new routes: GET /v1/agents/:agentId/deliveries/:deliveryId gives the state of a sent message and the turn's answer, POST /v1/agents/:agentId/settings changes the model, reasoning effort, provider or provider account, and POST /v1/agents/:agentId/context/clear starts a new chat, the same as New chat in the agent's settings. A change that the agent's state or the listed models refuse answers 409 with the reason.

Improved

Show a honeycomb of animated agent faces on the website, behind "Meet your first teammate" at the end of the home, news, guide, article and changelog pages. A blur under the text keeps it easy to read. Each face blinks, and the face under the pointer looks at it.
Message search in a chat and in the global search is faster on a long chat history. It reads the stored messages one time per search, not two times.
Published sites open faster. The site server now reads the block state and the site address list at the same time. Before, it read them one after the other on each request.
The app demo on the home page and the web client download only the language that they show. Before, they downloaded all ten languages, about 1 MB, before they started.
The home page loads less code from other pages of the site, so it shows sooner on a slow connection.
The Marketplace opens as a page of the window, as Schedule, Usage and Settings do. Its title is in the top row and the page rail shows that it is open. Escape goes back one view, or closes the page.
The buttons of the page rail show their names in a tooltip at their right. The tooltip opens and closes with the motion of a submenu, and it also opens when you move to a button with the keyboard.
When you move the pointer along the page rail, one tooltip moves from button to button, as a submenu does. The usage card of a provider moves in the same way.
In the web app, Settings and Server settings now open as pages of the window, the same as in the desktop app. The page rail marks the page that is open.
The web app page rail now has Schedule. It shows the routine calendar of the connected server.
The Help menu of the web app now shows the newest release notes.
Delete the Slack, Discord or Telegram Orchestrator when you disconnect its connector. Before, the orchestrator stayed in the sidebar with no connector. The Telegram Orchestrator goes with the last linked chat. When no agent is left in the Integrations section, the section goes too. The disconnect dialog tells you this before you confirm.
On the website, some agents in the agent network are larger leads with a team of small agents around them.
Shorten button, tab, menu and badge labels in all 9 translations, so that each label fits on one line. For example, "Retry" is now "Wiederholen" in German and "Ponów" in Polish.
Translate the text that showed in English in all 9 translations, such as the task panel and settings transfer.
Admins can now change a member's role, remove a member and end a member's session, as the owner can. Nobody can change or remove the owner.
The web app downloads less code before it shows the sign-in page and the workspace. It loads math formulas, Schedule, routines and server setup only when you open them.
A streamed reply in the web app uses less processor time in a long conversation. When you go back to the web app window, it no longer reads the whole server again if it read it in the last 30 seconds.
The web app uploads and shows large attachments faster.
On Windows and Linux, the main window no longer shows the menu bar under the title bar. Menu shortcuts work as before, and the Alt key shows the menu bar.
Set what an agent may do with one Permissions setting in its agent settings: Workspace only, Ask first or Full access. Ask first gives the agent full reach, and it asks you before it uses an OpenBot tool or publishes a site, and when its provider asks for approval. A new Claude agent shows Workspace only and Full access, because Claude does not ask for approval with full reach. On a joined server, an owner or admin can change the Permissions of its agents.
Full access agents now publish, update and delete public sites and use built-in tools without asking, as Turbo mode did before. Before, only Turbo mode let an agent do this without asking.
A Claude agent with Auto approve off shows Ask first. It asks before it uses an OpenBot tool or publishes a site, as before, but Claude's own tools run without asking. Select Full access or Workspace only to change it.
"Always allow" on an approval card shows only at Ask first, and moves the agent to Full access.
Make agent avatars do less rendering work when the pointer is outside the window. They no longer force a page layout on every animation frame.
Open the app about 1 to 2 seconds faster. At start, OpenBot no longer reads and checks every installed provider CLI again when no file of it changed since the last check passed. A changed, added or removed file still gets the full check.
Make long chats scroll and open with less work. A message that is not streaming no longer makes the app recalculate the page style when it comes into view.

Removed

The Usage page no longer shows the plan limits of each provider. The limit rings on the page rail still show them.
Remove Turbo mode from General settings and the Auto approve switch from the model menu. Each agent keeps what it could do: an agent with Turbo mode on before the upgrade has Full access.

Fixed

Messages save faster in a large history. Before, each save read all the saved command records, so it became slower over time.
Show no "Provider error" message when Grok cannot read its usage. Before, the message came back on each usage refresh. The usage still shows as unavailable.
Keep queue updates fast when an agent has received many messages. Before, each queue change searched all mailbox messages once for each message of the agent, so the app became slower as the history grew.
The daily billing snapshot reads only open subscriptions from Stripe. Before, it also read every canceled subscription, and one subscription with bad data stopped all snapshots. Now the snapshot counts such a subscription in skipped_subscriptions and leaves it out.
A chat update finds the attachments of each changed message from an index. Before, each changed message read every attachment in the database, so a large mailbox sync or a long chat slowed OpenBot.
Long chats stream replies more smoothly. Before, a closed chat search read the full chat again for each streamed part of a reply.
Keep the browser client fast while an agent writes a reply. Before, each part of the reply made every message in the chat again, so long chats became slow.
Send each sign-out and account change to connected apps once. Before, the account service could send the same change two times, and a slow connection service could stop the cloud server check for minutes.
Read Claude session history faster at start and after a provider reconnect. Before, OpenBot wrote each message of the selected conversation to its history index in a separate transaction, so a long session took seconds to read.
The Y axis of the daily usage chart shows each amount at its grid line. Before, the top amount was below its line and the zero amount did not show.
A provider with no known cost shows its tokens only. Before, its row said "Unavailable of known cost".
Start faster when agents have a long chat history. Before, each start and each provider reconnect imported the full provider history of all agent sessions again. Now OpenBot imports only new or changed turns, and does not read a replaced session again after one complete import.
Stop the database growth from agent list writes. Before, each message and each turn end added a full copy of the agent list to openbot.db, and the old copies stayed. The first write after the update removes them. Model changes that one agent made to another agent stay recorded.
The routine flow check after each turn and queue change reads only the runs it can move. Before, it read every routine run and flow step of the last 7 days each time, so a routine that runs often slowed OpenBot.
The home page title, text and buttons no longer move down when the font loads on a slow connection.
A reply that streams finds the earlier parts of its message from an index. Before, each update of a streaming reply read every earlier event of the chat, so a long chat streamed more slowly over time.
The server menu no longer shows Usage and Schedule for each server. The page rail opens them, as the 0.42.0 notes say.
The global search command for the routine calendar is now named Open schedule, the same as the Schedule page.
Keep only the provider CLI version in use and one older version that works. Before, OpenBot kept a copy of each update for 30 days, and this could use more than 5 GB. OpenBot now removes old versions within a few hours while it runs, not only when it starts.
Show provider runtimes as a separate part of the Storage bar. Before, they were part of "App data", which was then almost all of the storage.
Keep the usage rings of the side rail and their usage card on screen while models load. Before, the rings went away and came back each time a provider finished its check, so the card blinked.
The Schedule now loads when you change to a different server while it is open. Before, it stayed on loading until you closed it and opened it again.
On a phone, the web app now shows Usage, Schedule and the settings pages when you open them from the rail. The settings page puts its sections above the panel.
The side panel tabs are one tab stop. The arrow keys, Home and End select a tab, and Delete closes it. Screen readers announce them as tabs.
When Settings or Server settings opens, the focus moves to the page heading. When Settings closes, the focus goes back to the Settings button.
The settings search tells screen readers how many settings match.
Show local workspace and shared HTML files as pages in the sidebar when their links open. Keep the source view and external open option, using the same isolated page preview as HTML attachments.
The pstack skill now shows in the Marketplace. The 0.38.1 notes added it, but the Marketplace did not have it.
Grok agents answer again with Grok CLI 1.0.60. Before, every new Grok session failed with Invalid params: additionalDirectories is not supported, also when an agent switched to Grok.
The log now records a message that did not start because the provider refused the session or the turn.
Gemini connects on Windows when its server takes longer than 30 s to start. After a sign-in, a key change, or a start that did not finish in time, OpenBot waits up to 3 minutes for the next start. Before, OpenBot stopped the server while it unpacked, and each try failed with "ACP initialization timed out." The error now says which step did not finish: the start or the model list. The log records when each provider process starts, first answers, and ends.
On the website, the agent network fades out at the bottom of its section. Before, the agents were cut off at the bottom edge.
Fix translations with the wrong meaning, a mixed term or a mixed form of address. Before, the Polish mobile app showed some Russian text.
Show an ellipsis at the end of a long label in most badges and menu items. Before, the text stopped at the edge with no mark. Tab labels no longer go onto a second line.
Keep the side panel tabs when you go back to a chat from another page. Before, the open pane showed with no tabs.
Show what the agent works on in Computer Use screenshots and remote desktop on hosted and self-hosted servers and in the Docker image. Before, they showed the OpenBot window. Agents now get their own screen, and OpenBot moves to a hidden one.
Use the GPU again on Linux. Before, OpenBot always rendered on the CPU on Linux, so dialogs and panels with a blurred background were slow on computers whose GPU works. If the GPU process crashes, or a start stops before it is complete, later starts use software rendering, as before, until an update with a new Chromium version. A server (OPENBOT_SERVER=1 or a hosted server) always uses software rendering. Set OPENBOT_DISABLE_GPU=1 to always use software rendering, or OPENBOT_DISABLE_GPU=0 to try the GPU again.
An agent now sees why a browser file upload failed, for example a selector that matches two file inputs or a file that does not exist, and can try again with a better target. Before, a Codex agent saw only "dynamic tool request failed", and OpenBot showed a "Provider error" message.
When the server reads its member list, it now disconnects a member that was removed or got another role. Before, the open session could continue with the old role until it ended.
Update the Claude plan limits while an agent works. Before, the limits changed only on the next refresh.
Read the Claude plan limits faster when the computer has a large Claude history. Before, the read could time out.
Keep a record when an agent delete fails. Before, a host on Windows kept no record of the step that failed. Now logs/agents/agents.log shows the step and its error code, with no file paths.
Muse turns no longer fail when the workspace appears more than once in the writable roots.
Keep MCP servers such as Vercel signed in after their access token expires. Before, OpenBot did not ask for a refresh token, so the server refused agents one hour after each sign-in.
Show "Sign-in needed" for an MCP server whose access token expired and cannot be refreshed. Before, the row also showed "Signed in".
In the desktop app, the name of a rail button now moves to the next button when the pointer moves down the rail. Before, the name closed between two buttons and opened again.
Keep a routine open in agent settings while the agent works. Before, each new turn or status update of the agent closed the routine and went back to the agent settings, and the edit in progress was lost. Memories, Skills, Tables and Files in agent settings also stay open now.
Fix spreadsheet previews for January and February 1900. Keep Excel's February 29, 1900 compatibility date and the correct time in date-and-time cells.
An OpenCode agent no longer fails each message and routine run with "The request contains invalid parameters" after a model change or in a long chat. OpenBot now starts a new OpenCode session with the chat history and sends the message again.
A Muse agent no longer fails each message with "MCP startup audit failed; MCP is disabled for this runtime" after an MCP server, such as Computer Use, did not start once. OpenBot now starts a new Muse session with the chat history and sends the message again.
GIF and AVIF attachments keep their image format when sent through ACP, Muse and Pi.
Keep quoted currency and text prefixes in spreadsheet number previews.
The Antigravity model picker shows each Gemini model one time, with its efforts in the Effort menu. The model and effort that you select now apply to the session. An agent saved on a Gemini model with an effort in its name keeps that model and effort.
A message that you send while Codex compacts the context now waits in the queue and runs after the compaction. Before, Codex refused it and the message failed. When a compaction takes too long, OpenBot now asks Codex to stop it before the queue continues. Thanks to @Johnny-xuan for the report and the first fix (#1766).

Security

Open an attachment with its default app only when it is an image, audio, video, document or plain text file. Other files, such as scripts, web pages, programs and old Office files that can hold macros, show in the file manager. Before, Open on a file from a remote server or from a team member could run it.
Remove characters that are not permitted in a file name from the name of a file that a remote server sends, before OpenBot keeps a copy to open it.
Limit the downloads of a page in the browser to 4 at one time and to 2 GB each. Before, a page could download files until the disk was full.
Make the remote connection service safer. It now closes a connection that does not sign in within 30 seconds, and it gives the same relay credentials again when a device asks for new ones too frequently. Before, an idle connection stayed open for up to one day, and each request made new credentials.
Stop the relay server from sending data to more reserved network ranges, such as NAT64 and benchmark ranges.
Remove more types of secrets from logs, diagnostics and exports: JSON Web Tokens, private keys, every cookie in a cookie header, credential values in web addresses, and Google, GitLab, npm, Hugging Face, Groq, AWS session and Stripe keys. Before, a secret of these types from a provider message could stay in a log.
Disable and remove members correctly when an admin uses a team app on a server with an OpenBot account. Before, the change was only in the team file of the server: the account kept the membership, and the member stayed connected. Now the account changes, and the server closes each session of that member. This also applies to a change from the desktop app.
Count failed sign-in attempts for each connected device. Before, all devices that connect through the account service used one count, so one member could stop password sign-in to a user name for all other members.
Count sign-in attempts for a user name with spaces around it as attempts for that user name.
Send a server invitation email only for an open invitation that an owner or admin of that server made for that address. Before, any signed-in account could send an OpenBot invitation email to any address.
Show "OpenBot user" in the skill and agent marketplaces for a creator with no display name, and do not search creators by email. Before, the marketplaces showed the email address of such a creator.

Oct 16, 2026

Yaji 0.42.0

New

The website has a new section, "Your agents talk to each other". In an animation, Ada, Linus and Grace pass a task through OpenBot. Then a network that is different at each visit grows until it fills the screen. With reduced motion, the section shows the complete network.
A Help button in the page rail opens a menu with the newest release notes, the full changelog, the guides and Send feedback.

Improved

The license names Synthetify Labs spółka z ograniczoną odpowiedzialnością as the licensor. For a commercial license, write to licenses@openbot.run. The macOS and Windows app copyright also names Synthetify Labs.
The desktop window has a page rail at the left and a darker frame. The sidebar, the chat and the side panel sit in one card. The rail opens Agents, Routines, Schedule, Usage, Marketplace and Settings, and it shows the plan limits of each provider as rings.
Settings is now a page of the window, not a dialog, with a search for settings. Server settings is a separate page: open it from the server menu. When you select a different server, the server settings page shows that server.
The server menu in the top row is now the default way to switch servers. If you chose the server rail before, it stays. To use the rail, open the server menu and select Server layout.
The chat header has one panel button. It opens the side panel, which shows Browser, Files, Tasks and Agent settings as tabs. Use + to add a tab that is not open, and Maximize to give the panel the full width. When the tab labels do not fit, the tabs that are not selected show only their icon. A pane does not repeat the title and close button of its tab, and its content fades at the scroll edges. The Tasks tab shows closed tasks below the open ones, grouped by day, so you can open and reopen them.
The account box at the bottom of the sidebar is removed. Sign out is now in Settings > Profile, and Billing stays in Settings. When an update is available, the update notice shows at the bottom of the sidebar in the colors of the old account box.
The server menu no longer lists Marketplace, Usage and Routines, because the page rail opens them.
The web app has the same page rail as the desktop app. The switch over the chat list and the account box are removed. Settings on the web holds Billing and Sign out.
The routine calendar is now named Schedule, the same as its rail button. The Routines button of the rail shows the routines list in the sidebar.

Fixed

Keep message sends and queue status changes fast when the chat history is long. Before, each change wrote all mailbox messages again, so the app became slower as the history grew.

Oct 16, 2026

Yaji 0.41.0

New

Add a contact form at openbot.run/contact. The Contact buttons on the website open it, and the message goes to the OpenBot team by email.
A new customer now gets a "Welcome to OpenBot!" email when the plan of their first cloud server starts. The email shows the plan, the server name, and a link to the web app.
An agent can work on tasks at the same time as its chat. Start a task with "New task" in the chat menu, or move a queued message to a task with "Run as task". The agent can also start a task itself.
Each task has its own conversation, progress, stop button and result. Open a task from its chip below the chat, or from the chat menu. Closed tasks stay under "Earlier tasks", where you can open them again.
Two tasks of an agent run at the same time. Change this from 1 to 4 in the agent's settings, under Advanced, "Tasks at once". Other tasks wait, and the task shows why it waits.
A custom ACP agent runs one turn at a time: its tasks and its chat wait for each other, and the settings show this.
When a task fails, its panel shows the error, also after a restart. The main chat does not.
Tasks run without Computer Use, so that only the agent's chat controls the screen. Each task has its own browser tabs.
You cannot change an agent's provider while its tasks run. After the change, the tasks continue on the new provider.
The transfer file keeps the "Tasks at once" setting of each agent.
Add an "OpenBot for Teams" section to the website. It shows managed servers, team environments, custom integrations and launch support, and its Contact us button opens the contact form.
Add a For Teams page to the website, at /teams. It shows the three setups (local, managed server and custom deployment), what a server adds, the steps of a custom deployment, and who it is for. The header and the Teams section on the home page open it, and its buttons open the contact form.

Improved

A full refund of the current payment for a cloud server now ends its plan at once. The server stops and keeps its data. Before, the plan continued and renewed.
The chat header has one menu for the computer, remote control, files, agent settings and tasks.
The log records when a provider starts and stops connecting, so a report of a provider that stays on "Connecting" can show the cause.

Fixed

Show why a Telegram connect failed on a joined server. Before, the error said only "Request failed.".
Stop the Telegram connect dialog from waiting for Telegram after a connect fails.
A hosted server could start with "OpenBot database failed its integrity check." after it stopped. The host saved the disk while OpenBot wrote its database, and kept a damaged copy. A hosted server now stops OpenBot and closes the database before the host saves the disk. A server gets this change when it updates and starts again.
A server that got the stop signal (for example, from systemctl stop or docker stop) ended at once. It now stops its services and closes the database, as the desktop app does when it quits.
The integrity check error is now in your language, and the log shows which parts of the database are damaged.
Muse shows "Connected" after you sign in with your Meta account or save a Meta API key. Before, Muse always showed "Muse authentication is not verified", also after a successful sign-in. When Muse has no credential, it now asks you to sign in.
Open an agent with very long messages from a remote device. Before, the connection to the server dropped and reconnected again and again. The server now sends fewer messages at a time, and you load the rest when you scroll up.
Show an error when a server reply is too large to send to a remote device. Before, the server closed the connection.
Keep the numbers in a Markdown list in order when you add an item between existing items in the message box. Before, the new item and the item below it had the same number.
A server signs in to OpenBot at start when a slow database check holds the start. Before, the sign-in failed and the server went online about one minute later.
Start OpenBot faster after a clean stop when the database is large. Before, each start read the full database file to check it, which took minutes on a large file. The check now runs only after a stop that did not close the database, or after a database upgrade.
Stop the database from growing each time a conversation is read. Before, each read wrote the latest messages again when nothing had changed, so openbot.db could grow by gigabytes. An upgrade deletes the repeated writes that follow each other in a chat, which are most of them, and makes the file smaller. On a large database, the first start after the upgrade takes longer.
A hosted server no longer stops soon after it starts while you read it in a focused window or tab. Before, it stopped about 15 minutes after a wake when you did not type or click, and showed as asleep. An app that has focus now keeps it on for at most 30 minutes after your last message, change or typing, or after you return to the app.
The desktop app and the web client wait while a hosted server starts. Before, they stopped waiting after 5 minutes and showed "did not start in time", also when the server came online a minute later. They now stop only when the machine runs and OpenBot on it does not connect in 15 minutes.
The desktop app and the web client show the start step and its time while a hosted server starts. Restart is not available while the machine starts.
The desktop app shows when a hosted server did not start, could not start, or has an ended plan.
When the cloud provider limits new starts, the account server tries the start again after a delay, and the desktop app and the web client say so. Before, they showed "Could not start".
A conversation and its unread count load from an index of the message order. Before, each load sorted the whole chat several times, so a client that opened many agents could stall OpenBot for seconds.

Oct 15, 2026

Yaji 0.40.0

New

Backup models for a usage limit. When the plan of an agent's model refuses a turn, the chat shows a card: send the queued message to another connected model now, or wait for the reset. By default the agent asks first. Agent settings › Backup models keeps an order; turn off "Ask before switching" and the agent moves to the first ready model in the order without a question. A note in the chat shows each switch, and the agent goes back to its own model when the limit ends. The agent keeps its workspace and thread.
Set a server timezone in Server settings → General. Agents use it for a new routine when the message does not give the sender's timezone. Before, a message that waited across a server restart created its routine in the timezone of the server's computer, often UTC.
Connect 1Password on a joined server, such as a server with no screen. In Settings › Connectors, an owner or admin of the server pastes a 1Password service account token. The server checks the token, keeps it in its own encrypted store, and keeps the connection after a restart. The page shows the vault names and the login count of that server. Disconnect acts on that server only. Both sides need this release; on an older server, 1Password does not show.
An agent routine on this computer can send each answer to a Slack, Discord or Telegram place. Choose the place in "Send answers to" in the routine editor.
The run history shows where each answer went. A delivery has its own status: a run that succeeded stays succeeded when its answer is not sent. A connection that does not run gets more attempts for one day. Use "Send again" for an answer that was not sent, or that OpenBot stopped during the send.
Agents can set what a routine does with runs that came due while OpenBot was off, and what it does at the account usage limit. An update that changes one of these settings keeps the other.
On a joined server, an owner or admin can see and change what a routine does with missed runs, and see the missed runs in the run history. The server must have this version of OpenBot.
Routines can run one time, at a local date and time. After the run, the routine stays as a paused record. You can cancel it before it runs. A run missed while OpenBot was closed runs one time at start, unless you set it to skip. A restart does not run it again. Agents use this schedule for reminders, not a yearly schedule. Only the desktop app on the computer that runs OpenBot shows one-time routines for now.
Several instances on one Docker host: openbot-instance backup, verify, restore and backups back up all data of an instance, check the backup in a temporary folder, and restore it with the same agents and conversations. To back up an instance each day, turn on its timer with systemctl enable --now openbot-instance-backup@<name>.timer.
Settings → Dynamic Island has a Servers list. Turn off a server to hide its agent activity and requests from the island. The server stays connected and its desktop notifications continue. A switch sets whether servers that you add later show in the island.
Add links to the OpenBot Discord community on openbot.run: a "Discord" item in the Resources column of the footer, and a Discord icon next to the GitHub and X icons.
Add discord.openbot.run, a short address that opens the OpenBot Discord community.
Agents fill forms in the built-in browser faster. They can fill up to 16 text fields in one step. They do not use this step for password, verification code, or card fields.
Send agent notifications to the OpenBot app on Android phones. The phone gets the same events as an iPhone: a finished task, an error, a question, an approval, or a step in the browser. The agent name and the text are sealed, so Google and the OpenBot account service cannot read them.
The website compares OpenBot with Instinct at /compare/instinct: where agents work, who holds your data, who approves actions, price, and where you can use each.

Improved

A self-hosted server and the Docker container use the memory guards of a hosted server. When free memory is low, new turns wait in the queue and the browser opens no new tab. A fixed number of turns run at the same time, from the memory limit of the service or the container. Before, these servers started every turn and could run out of memory.
openbot status on a self-hosted server or a container shows the memory limit, the free memory and the turn limit.
Admins of a hosted server can now restart it from the server menu. Before, only the owner could restart it.
The model picker in a conversation shows only the providers that are ready to use. Select + to connect another provider in Settings.
A disputed payment for a hosted server ends its plan at once. The server stops and keeps its disk. A refund keeps the plan and the server running.
After you delete an agent, a notice with an Undo button shows for 8 seconds. Press Undo to get the agent back in the same place. OpenBot deletes the agent only when the time ends. The time stops while the pointer is on the notice.
The "Contact" button in the openbot.run header is now "Community". It opens the OpenBot Discord community instead of the OpenBot account on X.
Docker and hosted servers now include bubblewrap. Codex recommends the system copy instead of its bundled copy.
Docker documentation explains browser access and how to check a new release before upgrading a self-hosted server. It includes the release metadata source, deployment-file review and cold backup steps.
The add server dialog now asks "Need more power or a custom solution?" in its footer, in all languages.
"Contact us" in the add server dialog opens an email with a subject and a short message to complete.
A phone gets no notification while you use the desktop app, also for a question, an approval, or a step in the browser. The desktop app shows these itself.

Fixed

An agent that sent many messages to teammates made each chat open and each finished turn slower, because OpenBot searched all saved message keys again for each message. It now reads them once.
File, agent, skill and MCP chips in a message are centred on their line again. Before, their icons sat low, the file type badge went out of its box, the agent face was clipped on the right, and a file chip could cut a short name such as provider-session.log with an ellipsis.
Skill and MCP chips in a message and in the message box have icons of the same size and line weight as file chips. Before, their icons were larger and heavier.
Telegram messages and captions now preserve repeated spaces and tabs. Previously, code indentation and tab-separated text lost their spacing.
Show Markdown block quotes as quotes in Slack, including quotes with up to three leading spaces. Before, the message showed the > mark as text and no quote.
Telegram animations are now imported once. Previously, the same file could appear twice as an attachment.
Show email link labels and addresses as readable text in Slack messages and conversation history. Before, these links showed raw Slack markup.
Show Slack group mention labels in incoming messages and history. Before, these mentions showed raw Slack markup.
Keep underscores inside words with Unicode letters, numbers, or combining marks in Telegram messages. These underscores no longer turn parts of the word into italic text.
Emoji and other supplementary Unicode characters stay correct when a long Slack or Discord message splits into posts. Before, a split could break such a character.
Discord author and channel names can recover after a failed lookup. Before, later lookups kept showing the ID.
Slack author and channel names now recover after a temporary lookup failure. Before, the names stayed as IDs until the connection restarted.
Gemini no longer fills the disk on Windows. Each start of the Gemini server unpacks about 310 MB into the temporary folder, and a start that was stopped or timed out left that copy behind. Now each Gemini process on Windows uses its own folder in the OpenBot data folder, and OpenBot removes it when the process ends and at each launch.
On Windows, Gemini does not start when the disk has less than 1 GB free, and says so.
After 3 failed starts in a row, OpenBot stops starting a provider by itself and shows why on the provider row. Refresh the providers to try again.
On Windows, to get back the disk space that Gemini used in earlier versions, quit OpenBot and delete the _MEI folders in %TEMP%.
On Linux, the invitation page now offers the Linux installer. Before, Download OpenBot gave the macOS installer.
A self-hosted server now starts again when sudo openbot update fails or is stopped. Before, the server stayed stopped until an operator started it.
A browser live view that moves to a new connection now stops the previous view. Before, the previous view kept the screencast and the page size active.
Many large chats no longer stop the app for a second or more when their history gets too large to keep in memory. Before, OpenBot measured all chats again for each old message that it removed.
A published site now sends a folder address without a final /, such as /guide, to /guide/. Before, the page opened at /guide, and its relative styles, images, scripts and links did not load.
Slack, Discord, Telegram and webhook connectors no longer stay at Connecting when Signal accepts the connection but does not answer. After 15 seconds, OpenBot closes that connection and tries again.
After a Codex protocol error, OpenBot now stops the old Codex process with SIGKILL when it does not stop within 2 seconds. Before, the old process could continue to run next to its replacement.
In a narrow settings column, the ACP registry actions move under the agent and wrap. Before, long translated labels hid the agent name, covered its version, and cut off Remove runtime.
Stopping Pi or Muse while a chat opens now also stops that chat's provider process. Before, the process could stay open after the provider stopped.
Record the correct owner when two agents create shared tables at the same time. Before, both tables could go to the first agent, and the other agent could not remove its own table. A table that the user makes outside OpenBot also stays without an owner now.
A transfer to another computer keeps executable workspace and shared files executable. Before, an imported script failed to run by its path.
ACP registry settings now keep a gap between the search and agent ID fields and their buttons, and between Install and Cancel. Before, these controls touched.
A transfer export stops with an error when it cannot read a workspace, skill or shared folder. Before, it left out the folder's files and reported a complete export.
Agents can now write, search, and delete full-text (FTS5) and R*Tree tables in the shared database. Before, SQLite refused every use of these tables after CREATE, refused to create an R*Tree table, and the table could not be removed. While such a table existed, the shared tables list also showed no row counts.
An agent can read the result page after it submits a form in the browser. Before, the click failed and the tab showed an empty page.
An agent that answers in Slack, Discord or Telegram no longer stops answering after a provider error. Before, a failed Muse turn start, a provider that exited, a turn that ended before its start was confirmed, or a turn start with no confirmation could hold every later message with "OpenBot is working on another request" until OpenBot restarted.
An agent in a Slack, Discord or Telegram conversation no longer waits for an answer to a question card that nobody there can see. It asks the question in its answer instead.
An approval request that did not reach Slack, Discord or Telegram is posted again when the conversation is connected. Before, the turn waited with no notice in the conversation.
The "Working" and "Waiting" posts in Slack, Discord and Telegram now end after a restart, a provider crash or a lost connection. Before, they stayed with a Stop button that did nothing. When an answer cannot be posted, the conversation now says that OpenBot could not finish the request.
Slack and Discord messages no longer wait minutes to start when a channel or user name lookup is slow.
A hosted server that wakes now handles the kept Slack, Discord and Telegram events of each workspace or chat separately. Before, one connection that could not start held back the events of all other connections, and each of its events waited 60 seconds.
A Telegram chat that you link while OpenBot connects to Signal again now shows in OpenBot. Before, the link could be lost, and you had to link the chat again.
An agent that answers in Discord, Slack or Telegram no longer stops answering after a restart. Before, a restart could make one of its conversations the agent's own chat. Then, after a routine ran there, every new message waited with "OpenBot is working on another request" until OpenBot restarted. A restart no longer does this. An agent that is already in this state keeps its chat and its history, and answers again.
The chat of such an agent shows its new messages in the app again.
Stop the Windows download page and the OpenBot 101 guide from saying that the installer is not code-signed. They now say that SYNTHETIFY LABS signs it, and tell you how to continue when SmartScreen still warns about a new release.
You can name a sidebar section with a Japanese, Chinese or Korean input method. Before, the Enter that confirms a conversion saved the section, and Escape closed the editor and lost the name.
Start custom agents that need Node or another tool from your shell when you open OpenBot from Finder, the Dock, or a login item. Before, they stopped with exit code 127. Check agent now uses the same PATH, and a PATH that you saved on the agent still wins.
The Model field in agent settings, first-agent setup and the Slack and Telegram agent forms uses the full width of its row in narrow windows. Before, in windows 720 pixels wide or less, the field stayed at 148 pixels and cut the model name.
A provider CLI that cannot update or set up now shows "Update failed" or "Setup failed". Before, each failure showed "Download failed", also when the download was complete and the new CLI did not start.
The provider log (logs/providers/providers.log) now records the provider, the version and the step where a CLI update failed (download, verify, install or activate). Before, the line did not name the step or the version, and it could hold your home folder path.
Connect for Muse opens the Meta sign-in page. Before, no page opened and the sign-in could not finish.
A failed provider CLI update no longer shows the system error and your home folder path in the provider row, the model picker, or the provider status that team admins read. The row now shows a short reason, such as "Try again.". The full cause stays in logs/providers/providers.log.
A failed provider CLI update now says "The update failed. Try again." Before, it said that the update was interrupted, also when nothing interrupted it.
Keep the account of an idle provider when its CLI update fails. Before, the provider showed that it needed a new sign-in, but the old CLI was still installed and signed in.
Do not show a "Provider error" when Codex uses its bundled bubblewrap. Before, Codex showed this error on each connect on Linux servers, but it worked.
Keep the Create agent button visible and clickable when the agent description is long or the window is small. Before, the suggestion cards covered the button.
A scheduled routine no longer starts after you pause, delete or reschedule it while another routine is starting. Before, OpenBot could still start the old occurrence. A routine that you edit at that time now uses your new instruction and settings.
Docker instance restore now verifies the backup before it stops the target or replaces its data. A backup with valid checksums but a damaged database could previously replace healthy data.
Auto approve and Turbo now accept Codex's request to run an MCP tool. Before, a new tool still asked each time. A Workspace only agent, a Computer Use tool, a tool set to ask every time, forms and sign-in requests still ask.

Oct 14, 2026

Yaji 0.39.1

Action needed after you upgrade

For an instance made before, run `openbot-instance setup`, then `openbot-instance upgrade <name>` once.

New

Add Vercel to the Apps catalog. Agents can list projects and deployments, read build and runtime logs, and find why a build failed after you sign in to your Vercel account.
Agents can capture the screen of the computer that runs them: one display, one window, or an area of a display. You approve each capture first, or Turbo approves it when it has a standing approval for the agent, and the image shows in the conversation. Computer Use in the agent settings turns screen capture on or off for each agent. On macOS, OpenBot asks for the Screen Recording permission and tells you how to turn it on.
Codex, Claude and Grok can have more than one account. In Settings › Providers, select Add account under a provider to sign in to one more account. Each account keeps its own sign-in and its own usage limits, and you can rename or remove an account there.
Each agent can use a different account of its provider. Select the account in the model picker. The agent keeps its account after a restart.
The Usage menu shows the limits of each account.
An owner or admin of a server can add, rename and remove the provider accounts of the server, and select the account of each agent, in the desktop app and in the web client. The sign-in page of a new account opens in the server's browser.
An instance of docker/instances now updates from the app, like a hosted server: Server settings › Updates and the update dialog pull the new image on the host and make the container again, with the same data. The container never gets the Docker socket: it writes an empty request file, and openbot-instance-update@<name>.path on the host runs the update as root. It installs only a newer version. See "Updates from the app" in docs/docker.md.

Improved

The actions of a message are now in one row under the message, with the time: Copy, React and Reply. Before, they were at the side of the message, and Copy was in a … menu that had only Copy.
Agents that answer in Slack, Discord and Telegram answer simple questions themselves. They ask a teammate only when the request needs the tools, accounts, files or memory of that teammate, or long work.

Oct 14, 2026

Yaji 0.39.0

New

The server connection notice has a Details button. It shows why the server is away (an update, a start, or a lost connection), each step until the workspace is back, and how long it has taken.
Slack, Discord and Telegram now show in Marketplace › Apps, next to GitHub, 1Password and Bitwarden.
Run several OpenBot Docker containers on one host, each with its own CPU, memory and disk limit, with docker/instances/openbot-instance. The containers cannot reach each other, the host or private networks, only the internet. See "Several instances on one host" in docs/docker.md.
Server settings > Transfer moves the agents of a server to another server, also between computers. Export one file from the old server and import it into the new one. The file holds the instructions, avatar, model, skills, routines and memories of each agent and, if you want, its workspace files, plus the MCP servers, custom model providers and sidebar sections of the server. It also works with a joined server when you are its owner or an admin, for files up to 100 MB. API keys, sign-ins and chat history are not in the file. After the import, OpenBot shows which providers to sign in to again.
An operator can sign a self-hosted server or Docker container in for somebody else with no email code: openbot login --claim, which reads the claim from standard input. The account service makes the one-time claim for the owner's email with the new SERVER_CLAIMS_ADMIN_TOKEN. The claim works once, for 15 minutes. See "Sign in with a server claim" in docs/self-hosted-server.md.

Improved

The "MCP" link in the Server settings header no longer shows a background on hover. Only the text gets brighter.
The openbot.run home page now shows the iPhone and Android apps next to macOS, Windows and Linux. The "Available on" line and the footer list them, and the download menu has a "Phone apps" group with links to TestFlight and Google Play.
On an iPhone or an Android phone, the main download button of the home page now opens the phone app's store page. Before, it offered the Mac or Linux installer, which a phone cannot open.
When you restart a hosted server, the dialog stays open and shows the restart: the server stops, starts, and is back online. You can hide the dialog; a notification then tells you when the server is back or did not start. If the server does not start, "Try again" restarts it from the dialog.
Connect on a provider row now uses the easiest sign-in for the place you are in. On this computer, it opens your browser. On a server, from the app, it opens the sign-in page in your browser and sends the result to the server. In a browser tab, it shows a code. The other way to sign in is in the row's ⋯ menu.
In a browser tab, a provider that only the OpenBot app can sign in now tells you to connect it in the app, with a link to get the app.
In a browser tab, Gemini's ⋯ menu now has "Paste the address", so you can connect it without the app.
The Mobile tab of the changelog page shows the stores that have each version of the mobile app: iOS, Android, or both. A store that does not have the version yet shows "In review". When the two stores released the version on different days, each store shows its own date.
On the routine canvas, each connection now ends in an arrow where it goes into the next card, so you can see the order in which a run goes from one agent to the next. The step number in the corner of each agent card is removed; the routine details panel still shows it.
On a joined server, 1Password and Bitwarden now show in Server settings › Connectors as "Your computer only". Before, they did not show.
When a joined server has a new OpenBot version, a dialog now shows the update steps: download, restart, and back online. One click starts the update. An admin can let the restart wait until agents are idle, restart at once, or cancel the restart. You can hide the dialog while the server updates; a notice keeps the status and opens the dialog again.
In Server settings > Updates, "Update when idle" and "Restart now" open the same dialog.
The update dialog opens each time you enter a server that has a new version. An update button next to the server name in the sidebar opens it again.
The openbot.run website now keeps the first visit of your browser (source, first page, day and campaign tags) in the browser. When you sign in on the web, or open the page after a desktop checkout, OpenBot saves it on your account for product analytics. It contains no referrer, IP address or identifier. PRIVACY.md describes it.

Fixed

In an MCP form, the › in the Server settings header had more space on its left than on its right. Now the space is the same on both sides.
When your ChatGPT sign-in expired, creating an agent showed "workspace routing discovery unauthorized (401)". Now OpenBot tells you to sign in again, and the model picker shows the sign-in action.
When an OpenCode model ends a turn with no answer, the chat now says so, and says when the model stopped after its reasoning. Before, it told you to check sign-in and billing, also for free models where neither is the cause.
When you select a server on another computer, OpenBot now connects to it at once. Before, if the connection to that computer stopped recently, the server could show "Connecting" for up to 5 minutes.
Keep a phone or another computer connected when it is slow to read updates, such as a phone in the background while an agent works. Before, it could disconnect with "The host event stream has a gap.", because this computer sent updates out of order or skipped one that it could not send.
The open waits of an agent, and their errors, now show in German, Spanish, French, Italian, Japanese, Polish, Brazilian Portuguese, Russian and Turkish. Before, they showed in English.
Some routine flow messages in French, Japanese, Polish, Brazilian Portuguese and Turkish, and some provider sign-in text in Italian, now show in the selected language. Before, they showed in English.
On a server, "Log in with code" in the ⋯ menu opened the sign-in page again. Now it shows a code.
On a server, the model picker's Connect and Claude's Sign in button did nothing. Now they start the sign-in.
On a server that you manage, the message box now shows the sign-in notice for a signed-out provider.
The app now connects again after a server on another computer restarts. Before, it could show "The remote request timed out" until you restarted the app. Retry now also replaces a connection that does not answer.
While a server on another computer is offline, a conversation shows one "Queue load failed" error, not a new one for each agent change. The same error again replaces the earlier one.
While a server on another computer restarts to install an update, the desktop app now shows "Updating" and the server name, not a connection error. The app now waits up to 30 minutes, not 10, before it treats that server as offline.
A server that goes offline while its update waits no longer shows "restarting to install an update". The app shows that text only when the server starts the update restart.
The chat card that suggests a connector now works for every connector, on a joined server and in the browser client. Before, it worked only for GitHub, and on a joined server it said that GitHub works only on this computer.
A GitHub sign-in that waits for the code now shows "Connecting". Before, it showed as not connected.
In Marketplace › Apps, Connect on the GitHub card now shows on a joined server that can run connectors. Before, it showed only when the server could manage MCP servers.
The Bitwarden card in Marketplace › Apps now shows the same lock mark as Server settings › Connectors.
On a joined server, the Routines view no longer says that the server does not support routine flows while OpenBot connects, reconnects or waits for the server to wake. The view stays empty until the server answers, as it does while routines load.

Oct 13, 2026

Yaji 0.38.2

New

Italian interface. Choose Italiano in Settings > General > Language, or keep System default on a computer that uses Italian. The web client uses it too.
A server sends iPhone notifications to its members when an agent finishes, stops with an error, or needs input, an approval or a step in the browser. It sends no answer or error to a member who uses a desktop app at that time. While the phone shows the Live Activity, the Live Activity shows the event in place of a notification, and lights up the screen and plays the sound when the phone chose that. It uses the choices of each phone, and seals the agent name and the text so that Apple and the OpenBot account service cannot read them. The server keeps the phones that get notifications, so they continue after a restart.
You can cancel a Gemini (or other browser) sign-in that waits for the browser. Before, the only way out was Restart, which opened another sign-in page.
While a provider connects, its row tells you to finish the sign-in in your browser.
Agents can now post to a connected Slack, Discord or Telegram place on their own initiative, from any chat or routine run. Ask an agent to "post the summary to #bot-inflow", or put it in a routine prompt. openbot.messaging_list_places lists the connections and their channels, and openbot.messaging_post posts the text, redacted. Before, an agent could only answer the conversation that addressed it.

Improved

When a queued message waits and no turn runs, the server log now names what holds the queue, so a queue that does not start can be diagnosed.
The sign-in code email leaves about one second sooner. Before, OpenBot waited for 10 replies from the mail server, one after another, before it showed the code field. Now it waits for 5.
Agents show a draft before they send a message for you, and they send it only after you approve it. A routine whose instruction asks for a send still sends. Before a payment, an agent reads back the item, price and total and asks you to confirm.
Agents give the answer first and do not invent facts or sources. An agent says done only after it checks the result.
When you say stop, an agent also stops each teammate that it gave work to, and tells you which teammates stopped.
Agents look for a teammate earlier, when a task or part of it fits another agent's title and description. Before, they looked mostly after a failure. This helps when you have many agents.
A message from one agent to another now contains all that the other agent needs, without your other private details. When the answer comes back, the agent checks it and tells you which facts have a source.
For a repeated or timed request, such as "every morning", an agent creates a routine or offers one.
Server settings > Connectors shows the open connector in the header, for example "Connectors › GitHub". Select "Connectors" to go back to the list. The "All connectors" button is removed.

Fixed

Open OpenBot from an invitation link on Windows. Before, Open OpenBot on the invitation page did nothing, because Windows had no handler for openbot:// links.
Show help on the invitation page after Open OpenBot is pressed, in case OpenBot does not open.
Agents now show in their sections when a server starts or reconnects. Before, the sidebar sometimes showed them ungrouped until you refreshed the app.
Opening or writing to a chat no longer shows "Read state failed" when the read mark names a message that is not saved yet or was removed, such as a queued message on a hosted server. Before, every open and every send showed the error again.
The Google sign-in for Gemini no longer stops when you return to OpenBot before you finish it in the browser on the first-run screen.
A failed Gemini sign-in now shows the reason, such as a timeout or Google's own message. Before, every failure showed "OpenBot could not connect Gemini. Try again."
When Google refuses a saved Gemini API key, OpenBot says so. Before, it asked you to sign in with Google, which could not help while the key was saved.
OpenBot no longer opens a Google sign-in page by itself when it checks Gemini with an expired sign-in. It shows "Sign in with Google to use Gemini." at once, without a long wait.
A Gemini chat whose session was lost now starts a new session. Before, every next message in that chat failed.
When the Gemini plan's quota is reached, the turn fails with a clear message. Before, the quota notice was saved as Gemini's answer.
Gemini no longer asks for approval of a tool after you press Stop.
Gemini no longer shows a "Provider error" notification for each line of its internal log, such as after a Stop or a failed request. The failed turn says what happened.
When Gemini fails a request, for example because the API key is not valid, the turn fails with a clear message. Before, the error was saved as Gemini's answer.
The Gemini row in Settings always has an API key button, so a saved key can be replaced or removed while Gemini shows Connected.
After you cancel a sign-in, the provider row shows again why it is not connected.
Provider rows in Settings show why a provider is not ready, or why its download failed, as text. Before, the reason was only in a tooltip.
A Connect that fails in Settings, the model picker or the chat composer now shows why.
The Gemini row and other downloaded providers show "Update available" in Settings, as the update notification does.
On the first-run screen, the main button downloads the selected provider when it is not downloaded, and says Download.
The Server settings header no longer moves by 1 pixel when you open or close an MCP form.

Oct 13, 2026

Yaji 0.38.1

Action needed after you upgrade

Put agents that work together in one sidebar section. Channels are no longer supported. OpenBot archives every channel when it starts and hides channels on the desktop, the web and the iPhone app. Agents in one section know each other and send work to each other first. Your channel history stays on your computer for now. The next release removes channels and their data.

New

A mention, a direct message, a thread reply or a button press in Slack, Discord or Telegram starts a hosted server that sleeps. Signal keeps that message, encrypted for the server, for up to 10 minutes, and the server answers it when it starts.
Dictation can use ElevenLabs Scribe v2 Realtime with your own API key. Select it in Settings → Voice and save the key. The text shows while you speak, and a finished sentence stays fixed. The audio goes to ElevenLabs only when you select this model.
The Routines canvas is now in the web app and on joined servers. Switch the sidebar to Routines to see an agent's routines, connect agents, edit their tasks, run a routine and ask the agent to edit the canvas. A member who can see an agent can change its canvas; only an owner or admin sees and changes webhook settings. The server must run this version of OpenBot.
Visual replies and HTML files show as pages in the web app, as on desktop. The page runs in a sandboxed frame and cannot read the web app.
Image attachments show as thumbnails in the web app, and audio and video attachments of 10 MB or less play in the chat after you select Play.
The iPhone app can search the files in the chats of this computer.
An agent can stop a request and wait for an event outside its turn, such as a build or a nightly run. The chat shows each open wait above the composer. The next routine run of the agent, or a local script that names the wait, continues the request with the original message and the agent's note. Use the ⋯ menu of a wait to show its request or cancel it.
Generate a permanent Mobile Connect link in Settings > Mobile Connect. You can show it again, copy it, replace it or revoke it. The link signs in a phone each time you use it, until you revoke it or sign out on this computer. Revoking the link also disconnects the phones that signed in with it.
Copy or cancel a one-time Mobile Connect code. A cancelled code cannot sign in a phone.
See in Connected devices whether each phone signed in with a one-time code or a permanent link.
Agents that answer in Slack, Discord or Telegram can save short facts about the workspace, such as who owns a project. Every new conversation in that Slack workspace, Discord server or Telegram chat sees them, in every channel. The agent's own chat does not. A fact from a direct message is not saved.
An agent that answers in Slack, Discord or Telegram can read more of the channel or thread when a message refers to something earlier.
The owner of a hosted server can restart it from the server menu, on desktop and on the web. Restart server is above Leave server and Delete server. The server stops and starts again, and the app connects again when the server is back.
Add the pstack skill to the Marketplace. Install it on the agents you choose, and it adds the 52 skills of the pstack plugin by Lauren Tan, each as its own skill. Type $pstack-poteto-mode, $pstack-why or another $pstack-<name> to use one. Removing pstack removes them all.

Improved

Desktop notifications now show the start of the agent's reply when it finishes, the question when it needs your input, and the command when it needs your approval. Before, they showed only "Finished working.", "Needs your input." or "Needs your approval.". Secrets in the text are hidden. To show only the fixed text, turn off "Show message preview" in Settings → Notifications.
An account can now have 50 hosted servers. Before, the limit was 3.
A hosted server with a Slack, Discord or Telegram connection now sleeps after 15 minutes with no use, as other hosted servers do. Before, a live connection kept it on all the time.
Long chats in the web app and on joined servers show the "Earlier" part of the day rail for messages that are not loaded yet, as on this computer. The server must run this version of OpenBot.
Agents in one sidebar section now know their section teammates and send work to them first. To make agents work together, put them in one section.
When you open an agent whose history is not loaded yet, the loading crew covers the chat until the history is ready. Before, a banner showed a Retry button that you could not use. A failed load still shows the notice with Retry.
The chat now keeps the answer of a quiet routine run. Each routine run row opens and closes the messages of its run. A quiet run that has finished starts closed, a run that reported something starts open, and a run that is in progress stays open until it finishes. A quiet answer does not count as unread and does not change the chat preview.
In Slack and Discord, a mention of a person shows that person's name to the agent, not an ID.

Fixed

Keep a link in your own message readable when you put the pointer on it. Before, the link text almost disappeared into the bubble. A link that shows its full address no longer opens a tooltip with the same address.
A remote connection that Signal refuses for a reason other than a rate limit, such as an invalid ticket, reconnects at once. Before, OpenBot waited 60 seconds as if Signal had limited its rate.
An image you paste or drop in an agent chat sends again after you had a channel open. Before, the closed channel deleted the attachment as it arrived, and sending failed with "Attachment draft no longer exists".
Claude background subagents keep running after the agent's turn ends. Before, OpenBot stopped the idle Claude process after 10 minutes and ended every subagent with "Stream closed", so long research had to start again.
The live view of a remote server's browser now stops when you hide the browser, change the tab, or close a sign-in page. Before, the server kept sending the old view.
A browser password request now shows the preview of its own page. Before, a slow preview of an earlier request could replace it.
When a server connection was lost while OpenBot checked it, for example when you came back to the app, the status showed a short countdown before it connected again. Now it connects again at once.
A Claude agent's answer after a background subagent or command ends now shows in the chat. Before, Claude wrote that answer, but OpenBot did not show it, because no message of yours had started the turn.
The chat over the Routines canvas shows the agent's answer once, not twice.
Start the channel task that has waited longest for the host first. Before, a busy channel could start its newer tasks again and again while an older task of another channel stayed queued.
Wait between connection attempts to a team server that is offline. Before, the app tried again about once each second, and Signal could then block the app for a minute.
The web client connects to a hosted server that starts within about 5 seconds after the server is ready. Before, it could wait up to 2 minutes between attempts.
While a hosted server starts, the web client shows why the last connection attempt failed, with its error code.
An owner or admin of a server sees its Slack, Discord and Telegram connectors in the web client's Server settings, as in the desktop app. Before, the web client showed only GitHub. A Slack or Discord install that the web client starts returns to the web client tab.
OpenBot uses Electron 44.7.0. This update has V8 engine fixes that can stop rare crashes of the app while it is idle on macOS.
A routine run that a local script or a webhook starts can now end quietly, as a scheduled run does. Before, OpenBot sent it as a Test run, so the agent always wrote a report. The run history now shows "Script" or "Webhook" for these runs, not "Manual".
A client now connects again when it asks to start a hosted server that is stopping. Before, it stopped the reconnects and showed the server as offline until the next status check.
Account sessions now show a hosted server with its name, as a hosted server. Before, the list showed it as "Desktop", so an owner could disconnect the server by mistake. Disconnect still signs the server out; a restart signs it in again.

Oct 12, 2026

Yaji 0.38.0

New

URL MCP servers have a "Sign-in app" section. Enter the client ID, an optional client secret and optional scopes of an OAuth app that you registered with the server's provider. OpenBot uses this app when the server cannot register OpenBot by itself. The section shows the exact callback URL with a copy button. On a joined server, the host keeps the app and signs in with its own browser.
On a joined server that supports it, you can now connect Codex, Claude, Gemini and Muse from your own computer. The sign-in page opens in your browser, and OpenBot sends the callback to the provider's CLI on the host.
On a joined server that supports it, Server settings > Connectors now shows the host's GitHub, Slack, Discord and Telegram connections. You can connect, disconnect and reconnect them from your computer. The pages open in your browser, and the host keeps every token.
When an agent on a joined server shows a sign-in link, for example for an MCP server of its CLI, the link opens in your browser and the sign-in finishes on the host.
The Marketplace can now install an app that signs in with a browser on a joined server that supports this relay. The browser client can also connect providers and install these apps on a joined server.
On a joined server that supports it, the sign-in page of a URL MCP server now opens in the browser of your own computer, not in a tab on the host. Your passwords, passkeys and existing sessions work. OpenBot takes the callback on your computer and sends it to the host, and the host keeps the grant.
When your computer cannot take the callback of a relayed sign-in, OpenBot asks you to paste the address that the browser shows at the end of the sign-in. The browser client always uses this step.

Improved

A hosted server installs a new release by itself when it is idle and no member used it for 5 minutes. Automatic updates are always on for a hosted server.
Show the words you dictate at the end of your draft, where they will go. Before, they showed as a separate grey paragraph under the draft and moved into it when the recording finished.
Make the voice recording bar shorter, so the message box has more space for text.
MCP sign-in now tries the same local callback port each time (28417), so the callback URL of a sign-in app stays the same. When the port is busy, OpenBot uses a random port.

Fixed

Show a moving bar and no percent while a server prepares an update or stages it. Before, the update notice showed "Downloading · 0%" for minutes while a cloud server waited for its disk after a resume.
A hosted server that had a new release ready took more than 3 minutes to wake. It now wakes in about 20 seconds, as it does without a release.
When an agent opened more websites in the built-in browser, or when you came back to the OpenBot window, the browser went back to the first website. It now keeps the website that you chose.
A paste into the live view of a remote host's browser now goes to the page. Before, the paste went to the last text selection in the window, such as the chat input, or it was lost.
A connection to a remote server that stayed open for 45 minutes asked the signalling service for new relay credentials many times a second, until the service closed the socket for a minute. The service could then also close the server owner's own host connection. OpenBot now asks once every 45 minutes, and restarts the network path only when it goes through a relay.
When you refuse a relayed sign-in on a joined server, the command-line tool on the server now gets the refusal and stops at once. Before, it waited until its own time limit.
The paste step of a relayed sign-in no longer says "this browser" in the desktop app.

Security

OpenBot keeps the client secret of a sign-in app in the encrypted sign-in store, shows only "Secret saved" after you save it, and removes it when you remove the server. Only an owner or admin can change a sign-in app on a joined server. A sign-in app is bound to the first sign-in server that accepts it, and OpenBot never sends it to a different one.
The host relays an agent's sign-in link only when the link is an https page with a state that returns to a loopback address, and only when a message of that agent holds that exact link. A Slack or Discord grant goes only to the server whose install your computer opened, and only once. No connector route carries a token, and OpenBot does not log sign-in codes or grants.
Only an owner or admin can start a relayed sign-in. Each relay completes once, only for the member who started it, and ends after 10 minutes. Your computer sends only the query of the callback. The host checks its state and never lets a client choose a host port or address. The page opens only when you click "Open sign-in page" in the dialog that shows its host name, and your computer refuses a page that has no state, so a host cannot use your browser for a grant you did not ask for. OpenBot does not log the sign-in code.

Oct 11, 2026

Yaji 0.37.0

New

A scheduled routine has the setting "If OpenBot was closed at the scheduled time". "Skip this run" is the default. "Run once when OpenBot opens" starts one run for the most recent missed time.
The run history shows the runs that OpenBot skipped while it was closed, as one row for each group, and marks a run that started more than one minute late.
Connect the phone app to a self-hosted or Docker server. Run openbot mobile-connect on the server (docker exec -it <container> openbot mobile-connect in Docker) and scan the QR code with the phone app. Before, only the desktop app could show a Mobile Connect code.
On this computer, you can log in to Grok, Cursor, Cline and Muse with a code, as you can for ChatGPT. On Linux, you can also log in to Claude with a code that you paste.
When a browser sign-in that you start does not finish after 25 seconds, or fails, the provider row and the composer show a "Log in with code" button.
On this computer, you can connect Gemini with a Gemini API key. Select Sign in on the Gemini row to save the key.
Settings > Voice shows the dictation models. You can download, cancel, and select a model there.
You can download Whisper large-v3 turbo (1 GB) as a second dictation model. It knows about 100 languages. With Whisper, you can set the dictation language, for example Hungarian, or keep Automatic. Whisper writes the text when you stop, not while you speak. Parakeet stays the default and finds the spoken language itself.

Improved

German, Spanish, French, Japanese, Brazilian Portuguese, Russian and Turkish now translate all interface text. Before, some screens, errors and status messages showed English text.

Fixed

Claude agents keep their prompt cache when their saved memories change. Before, each memory change, such as a new memory that the agent saved, made the next turn write the whole conversation to the cache again. The agent now gets the change with its next message.
Show "1 unread message" and "1 unread direct message" in the desktop sidebar. Before, these labels used the plural for one message.
Translate open notifications again when you change the language. Before, a provider update notification and the close button of each notification kept the previous language.
In the desktop app, an image that you paste or a file that you drop into a channel chat now appears as an attachment. Before, the channel chat ignored it, and only the attachment button worked.
An agent whose plan ran out of usage can now change to a model of another provider. Before, the change failed with "Could not change model" while the agent held a queued message, and the agent stayed on the spent plan until it reset. The queued message now runs on the new model.
Agents in a channel no longer read the thinking of other agents. Before, when one Claude agent in a channel had thinking text in its turn, Claude stopped every later assignment in that channel with a reasoning_extraction safeguard error until the channel summary covered the text. The channel still shows the thinking.
The chat input keeps focus while agents work in the background. Before, a browser popup that an agent opened, or a browser tab that an agent or a page closed, took focus from the chat input when the browser panel was open.
Keys that you type during a short provider restart or server reconnect are no longer lost.
A click into the chat input that closes a popover keeps the focus in the chat input. Before, the focus went back to the button that opened the popover.
Connect for Gemini now works on a Linux server that you joined, such as a Docker server. Before, the button did nothing. The server opens the Google page in its own browser, and you sign in through the live view of that page in the app or the web client. Only an owner or admin can do this.
A finished agent reply no longer stays cut off after about 220 characters until the window is reloaded. This happened most often in a desktop app connected to another computer or server. A streaming reply also no longer loses its middle part while other agents work, and the Dynamic Island no longer cuts a reply that it already shows.
After a joined server restarts, the first action on that server no longer fails with "Provider settings are not supported by this connection." Before, a second try was necessary.
Let users close the usage-limit notice above the message box. Keep the current notice dismissed when they return to the chat, and show a notice again for a new quota window.

Oct 10, 2026

Yaji 0.36.0

New

Polish interface. Choose Polski in Settings > General > Language, or keep System default on a computer that uses Polish. The web client uses it too.
A computer that has not signed in to OpenBot now opens on a language wheel, before sign-in. Turn it with the arrow keys, the scroll wheel, a drag or the arrow buttons; the screen changes to each language as you turn it. Continue saves the language. The wheel shows one time on each computer.
The diagnostics export now has the memory and CPU use of OpenBot and of each provider's processes, the app's responsiveness, database statement times and crash counts, for today and the previous day.
With analytics on, OpenBot sends one daily summary of its memory and CPU use, by computer class and provider. It does not send process names, command lines or paths.
Added a guide to run the Docker image as an AWS ECS service on EC2, with the seccomp profile, the data volume, ECS Exec sign-in and CloudWatch logs. AWS Fargate is not supported: it cannot keep the Electron sandbox on. See docs/aws-ecs.md.

Improved

The Docker image now works when /dev/shm is smaller than 1 GiB, for example on Kubernetes: Chromium then keeps its shared memory in /tmp. Before, Chromium could run out of shared memory.

Fixed

Completed the Brazilian Portuguese catalog: the 643 keys that showed English text — agent and provider settings, routines, webhooks, the routine canvas, billing, server states, the mobile app and the Marketplace — are now translated. bun run i18n:check reports pt at 6,676 of 6,676 keys.
A server that gets a stop signal while its window loads now quits. Before, it showed a "failed to start" box on its virtual display, and the box blocked the quit until the process was killed.

Oct 10, 2026

Yaji 0.35.2

Fixed

Agents on Muse models, such as Muse Spark 1.3, reply again. In 0.35.0 and 0.35.1, each message to these agents failed with "Recursive JSON schemas are not currently supported".
Keep the agent avatar inside the expanded Dynamic Island when the main window is zoomed in. Before, the avatar moved to the left of the panel and was cut off. The Computer Use rim and the permission help window also no longer change size with the main window zoom.

Oct 10, 2026

Yaji 0.35.1

Improved

The usage popover now shows each limit that a provider reports, such as the 5-hour limit and the weekly limit. Each limit has a bar, the percent left, and the time until it resets. Before, the popover showed only the limit with the least left.
A failed connection to a server, the web app or a hosted server now shows a specific reason, for example when the server is busy, an invitation was already used or a firewall blocks the direct connection. Before, most of these failures showed "The connection failed."
Error messages and error notifications show a short error code with a copy button. Send this code when you report a problem, so we can find the cause.

Fixed

A hosted server no longer stays at "Starting" with "Too many active remote connections." Before, the phones, browsers and desktops of an account could use all of its remote connections, and the server could not connect. Now only device connections count toward the account limit.
The server connection screen shows its errors in your language. Spanish now has the text of this screen.
The server menu on the sidebar title now shows all of its width while a server sleeps, wakes, or connects. Before, the loading screen of the server covered the part of the menu that went past the sidebar.
The web app connects again by itself when a server comes back after an update or a long restart. Before, the page could wait up to two minutes before it tried again, and you often had to reload it.
A device that was connected when its server restarted no longer stays at "Loading" until you reload.
A server that lost its remote connection for good no longer shows as online while no device can reach it. The server now reports an error, and a cloud server connects again by itself.

Oct 10, 2026

Yaji 0.35.0

New

An owner or admin can turn Computer Use and Local scripts on or off for an agent, and select what a message does while the agent works, from the OpenBot iPhone app.
On macOS, you can reply to an agent, answer its question and stop its work from the Dynamic Island. The main window stays where it is. "Open chat" still opens the conversation.
Add native Pi RPC and Muse MSP providers, managed runtime downloads, and host sign-in.
Add ACP provider settings with saved overrides, deferred changes, and reset.
Add ACP registry search, install, update, cancel, and runtime removal on the selected host.
Apps connected to a server that runs this version can manage its providers and ACP agents. Custom agents keep their IDs and sessions.
An owner or admin can sign in to an MCP server that uses OAuth on a joined server, from the desktop app or the web app. The sign-in page opens in a private tab of the host's browser and shows in the settings. The token stays on the host.
The file preview panel has a Copy button for text, code, Markdown and HTML files. It copies the full source of the file, also for Markdown and HTML that the panel shows rendered.
The file preview panel now highlights code files, such as JSON, TypeScript, Python, YAML and diff files, with the same colors as code in chat. Files larger than 200,000 characters, or with a line longer than 5,000 characters, stay as plain text.
The file preview panel has a wrap button for text files and source views. When it is on, long lines wrap to the panel width, so you do not have to scroll sideways. The app remembers your choice.
Agents can search saved memories and list them for maintenance. Essential memories use a fixed prompt budget.
You can choose which memories are always included, which need search, and which the agent can select.
Show selected release notes after a desktop update. Updates that skip versions show the selected notes together. Close the dialog to continue, or open it again from Settings > Updates. A new installation starts without the dialog.
Add Cobrain to the Apps catalog. Agents can load, search and save notes in your Cobrain memory after you sign in to your Cobrain account.

Improved

The agent settings panel shows the face, name, title and instructions at the top of the list, as the iPhone app does. You no longer open separate Profile and Instructions pages. Click the face to change the image, the generated face or the color.
The right panel cannot be narrower than its default width, because a narrower panel breaks its layout. A panel that you made narrower before opens at the default width.
The instructions field of an agent and of a channel has a fixed height. Long instructions scroll inside the field. Before, the field grew and had a resize handle that did not match the field.
Voice input now uses NVIDIA Parakeet TDT 0.6B v3 in place of Whisper. The words show in the message box while you speak, and the full text replaces them when you stop. Parakeet supports 25 European languages. Japanese, Chinese and Korean dictation is no longer available.
The first voice recording downloads a new model of approximately 670 MB. OpenBot deletes the old Whisper files.
While you speak, the microphone button changes into a recording bar with a live waveform. The waveform shows small dots in silence and taller bars when you speak. ✓ puts the text in the message box, × discards the recording, and the send arrow sends it. The message box no longer gets a yellow outline during a recording.
Local scripts can send messages, read pending questions and approvals, and submit responses. Agents with Local scripts already enabled get these operations after upgrade without a new opt-in. Browser takeovers and secret questions direct the user to OpenBot.
Owners and administrators can turn Local scripts on or off from a connected desktop or web client. Scripts run on the host, so a headless server needs no local screen.
Existing memories stay saved. During upgrade, recent manual memories receive priority for the essential prompt budget.
Remote memory selection uses bounded pages. Older hosts keep memory text editing without the new controls.
A scheduled routine run can now end without a message when the agent calls the new routine_no_update tool. Before, the run stayed quiet only when the answer was exactly [[no-update]], and an answer such as "Nothing new." sent a notification. The marker still works. A Test run still shows its answer.
Agents use less context. They load the details of most tools only when they need them. All tools stay available.
When an agent changes provider, it gets a shorter part of the conversation. It can still read the older messages of the same conversation. A context reset still hides earlier messages.
A change of the Claude model or reasoning effort no longer starts a new Claude session.
A ring near the message box shows how much context the agent uses. The conversation shows when the agent compacts its context. Claude values are estimates.

Fixed

A paste into a remote desktop now pastes your own clipboard on the host. Before, Cmd+V or Ctrl+V pasted the host's clipboard.
The embedded browser now removes the OpenBot and Electron user-agent tokens for sites such as X and LinkedIn, and sends basic client hints for its Chromium version and platform. Google account requests keep the compatibility token needed for sign-in.
The owner of a server can now remove it from the account with Remove server, in the server menu or in server settings, also when its computer is offline or gone. Before, an owner could not remove such a server, and it came back on each computer and after each reinstall. The removal applies to all members and all devices. Members lose access, and connected Slack, Discord, Telegram and webhook routes stop. If you add the server again, invite its members again and set up its connections again. Old webhook URLs do not work again. Files and chats on the server's computer stay.
Keep chat input focused while another agent uses the browser. Background browser actions could interrupt typing.
Keep chats at the latest message when switching agents. Measurements of routine cards and other messages could move the chat away from the bottom.
Custom ACP agents can recover from a closed session when you send a new message. Before, messages could repeatedly fail with Internal error. Startup errors now show provider details with secrets removed. Recovery keeps the conversation and does not repeat a failed prompt.
Codex now saves “Always allow” approvals for custom MCP server tools and applies saved approvals after a restart.
An agent whose provider says "You've reached your weekly usage limit" now waits for the limit to reset, as it does for "You've hit your weekly limit". Before, each queued message and routine run started a turn that failed.
Switch chats while OpenBot changes which queued message you edit, and each chat now keeps its own draft. Before, the first chat could get the other chat's draft and lose its own.
A voice message sent to one chat while you read another no longer scrolls the chat you read to its latest message.
Close the chat search or clear the query while a search is running, and a late result no longer opens an old match, possibly in another chat.
If a voice recording fails to start after the microphone opened, OpenBot now releases the microphone. Before, the microphone stayed in use.
A Stop that arrives after its turn ended no longer stops the next turn of the same agent.
A file or site folder whose name starts with two dots, such as ..notes.md, now opens or publishes from the workspace. Before, OpenBot refused it as outside the workspace.
If saving the preview of a sent message fails, the message is still delivered. Before, it stayed in the queue until the next restart or queue event.
A routine set to repeat every 58 or 59 minutes is now refused when you save it, like every other schedule under the 3-minute minimum. Before, it ran twice within 1 or 2 minutes around each hour. Routines already saved this way keep running and can still be edited.
The sidebar section name field keeps focus after you choose New section or Rename. Before, the field could lose focus and the name was lost.
Remote Control keeps its stream when another server changes, for example when it goes offline or you mute it. Before, the view changed to "Remote access expired".
Text you type in a direct message stays when someone on the server starts typing to an agent or connects. Before, the message box cleared and the conversation scrolled back.
Deleting a routine now also removes the work it queued and nobody started yet, in agent chats and in channels. Before, that work could still run after the routine was gone. Channel work that already started keeps running and can be stopped as usual. Deleting an agent routine still stops its active turn.
Files attached to a channel message are no longer lost when the channel is archived while the message is being sent. You can send them again.
Stopping or reassigning a channel task while it was being handed to a member no longer gets undone when that hand-off fails.
Show Gemini replies as they arrive while Antigravity background tasks run. Stop now ends the local turn, saves unfinished tool calls as interrupted, and restores the same provider session before the next message.
A queued message now appears at the point where you select Steer, after the commentary sent while it waited. Its position stays the same after a restart.
Show when a queued message is held for editing, including after restart, and add a Cancel edit button.
Keep pending messages visible when no agent turn is active, so users can edit or cancel them.
Two attachments whose names differ only in case, such as Report.txt and report.txt, now both arrive. Before, on macOS and Windows the second file overwrote the first, and the first attachment could not be opened.
A Claude turn that ends on an API error, such as an expired sign-in, a low credit balance or an overloaded service, now fails and shows the error. Before, the error text was saved as the agent's answer and the message counted as delivered.
When an agent cannot attach a file to its answer, it now gets the reason, for example a file outside its workspace and the shared directory, and can correct the call. Before, Codex agents got only dynamic tool request failed and retried the same call.
A Codex agent on a Plus or Pro plan now waits for its usage limit to reset. Before, the limit message offered more credits, so OpenBot took it for a spent balance and started each queued message and routine run, and each one failed.
Grok agents answer again after their Grok CLI restarts. Before, every message to the agent failed with Invalid params after an idle stop, an OpenBot restart or an update.
Fix remote reconnects that report too many active connections when an old client socket waits for its host. Concurrent reconnects and failed authentication no longer hold extra account slots. Other device sessions stay connected.
Wait before reconnecting when Signal limits connection traffic. Before, repeated retries could keep all devices on an account disconnected.
Gemini shows its sign-in and start errors. Before, it could show Ready when it could not run.
Restart works while a Gemini sign-in waits. Before, the old sign-in could stop the next attempt.
On Linux, Gemini sign-in opens in the desktop browser. When no browser opens, OpenBot shows the setup steps. Before, it waited until the sign-in expired.
Remote provider settings explain that Gemini sign-in opens a browser on the server's computer.
The message box and its queue stay below an approval request or a browser takeover. A message that you send while an agent waits for your approval now shows in the queue. Before, the message box went away, and a queued message was not visible until the approval was done.
When an agent refuses a request, OpenBot now shows the reason that the agent sends. Before, a Grok agent showed only Invalid params, for example when Grok refused a session resume after its CLI restarted.
Messages stay in the queue when you try to steer a provider that does not support it. Custom ACP agents no longer receive a second prompt that can interrupt the running task or lose the queued message.
Recover the affected remote peer when a host data channel closes. Other connected clients stay connected.
Report failed Live View connections and invalid frames as failures. Close failed views so they can be opened again.
A queued channel task now starts when the agent or normal work that held it becomes free, also when that work ended without a turn, for example when a message failed to start or was cancelled. Before, the task could stay queued until the next channel command or a restart of OpenBot.
A message from an agent with an empty reply id no longer stops the sender's and the recipient's conversations from loading or stalls the recipient's queue. Conversations that an earlier version saved with such a message load again.

Security

The Bun runtime that OpenBot downloads to start MCP servers is now version 1.4.3, which has security fixes.

Oct 9, 2026

Yaji 0.34.1

Improved

The mobile app card at the bottom of the sidebar now says "OpenBot for mobile", because the app is on iPhone and Android. "How to install" opens Settings > Mobile Connect, which shows the steps for both phones. If you closed the old iPhone card, the new card shows one more time.

Fixed

The message box stays at the bottom of the chat again. In a short conversation it moved up to sit directly below the last message.

Oct 9, 2026

Yaji 0.34.0

Action needed after you upgrade

Update the Bun development and build toolchain to 1.4.2 across desktop, mobile, and server workflows. Developers must install Bun 1.4.2 with the command in the README before they start development.

New

A Routines view in the sidebar. Open an agent to see its routines on a canvas: when each one runs next, its last runs, and what each agent received and answered in the last run.
Routine flows. On the canvas, connect one agent to another inside a routine, and the routine sends the first agent's answer on to the next agent when it runs. An agent with several inputs waits for all of them. Routine flows are available for agents on this computer.
Ask an agent to edit the canvas. The chat panel on the Routines view sends your request to the open agent, which can add or remove routine links, and create or change routines, with its tools.
Change an agent's task in a routine from the routine details panel. For the routine's own agent, this changes what the routine asks for.
Right-click the routine canvas to add an agent at that spot: one you already have, or a new one with its name, what it does, and its model.
Webhook routines show on the routine canvas. Their details set the webhook: copy the endpoint, regenerate the signing secret, and choose the event type and filters that start a run. Run now tests them.
All signed-in server members can check for releases and request an idle update on supported installations. Administrators keep control of cancellation, forced restarts, and automatic-update settings.
Turn built-in providers off or on in this computer's provider settings. OpenBot saves the choice, stops unused provider processes, skips their CLI checks, and removes their models from the picker. Change an agent's model before you turn its provider off.
Russian interface. Choose Русский in Settings > General > Language, or keep System default on a computer that uses Russian. The web client uses it too.
Add Spanish and German to the language settings. Translate desktop and web-client screens, including setup, agent controls, routines, files, connections, status messages and errors.
The server menu on the server rail and on the sidebar title now has "Leave server" for a server that you joined and "Delete server" for a server that you host. The desktop app and the web client both show these actions.
Update a hosted server, or a server that you installed on Linux, from Server Settings > Updates. Download, Update when idle, Restart now and automatic updates now work there as on a computer. Before, the server got a new release only when it started again.
Remove an old server that you own from the web side panel, even when its computer is offline. Removal ends remote access for all members and keeps files and chats on the computer. Paid hosted servers still use Billing.
Let a scheduled run of an agent routine stay silent when it has nothing to report. Ask the agent in the routine task to answer [[no-update]] in that case. When every answer of the run is only [[no-update]], the run posts no message, adds no unread message, does not change the chat preview and shows no notification. The run marker stays in the chat, and the run stays in the routine History. Test runs, and script or webhook runs, always show their result. The browser client, the phone, and a desktop connected to a remote server stay silent for a quiet run too. An older client shows the run as finished, as before.
Sign in to a custom Streamable HTTP MCP server in your browser, for servers such as Granola that use OAuth. Choose Sign in on the server's row or in its form. The panel shows that it waits for the browser, lets you cancel, and marks the row Signed in. Sign out deletes the stored sign-in.
The Custom tab of the model picker now has an "Add provider" button. It opens the Providers section of the server settings, where you add an endpoint. Before, the tab had no way to add one. The button shows on this computer, and on a host where you are an owner or admin.
The iPhone and Android apps have their own download pages, /download/ios and /download/android, with the system requirements, the install steps and how to connect the phone to your computer.
Telegram: add the OpenBot bot to a Telegram group or start a direct chat with it from Server settings → Connectors → Telegram. People mention the bot or reply to it, and the Telegram Orchestrator asks the right agent and posts the answer in the same reply chain, with approvals, stop and files as in Slack.
OpenBot sends the page's mouse cursor to the mobile app, which shows it on the live browser view.
A right-click from the mobile app opens the page's menu on the phone, not on the screen of the server. While the phone shows a tab, the page keeps one size, so it does not change shape when the OpenBot window changes size. A tab that an agent gave its own size keeps that size.
Choose how many memories each agent can keep, from 64 to 512, in Settings > General. A lower limit keeps all saved memories.

Improved

The Routines tab of the sidebar shows only agents. Channels have no routines, so they stay in the Agents tab. When you open the Routines tab, an open channel closes.
Development commands can start without shared secrets and load the encrypted shared development settings when the developer has the development key.
Local development identity and overrides persist in ignored state, and the APNs helper writes to that state.
Encrypted development settings use .env.dev and the shell key DOTENV_PRIVATE_KEY_DEV; production settings remain unchanged.
The update resets old generated .env.dev settings. Existing identity keys and overrides in .openbot/dev-state.json remain unchanged.
Development commands keep account and decryption secrets out of browser and tunnel processes, and secret-dependent E2E commands check required keys before changing test data.
New hosted servers start from a snapshot built from the latest completed desktop release. Releases now build and select the production boat snapshot automatically. Failed snapshot builds keep the previous template active.
Marketplace app listings in Russian. The tagline, description, server description and example requests of each catalog app now follow the interface language. Before, they were always in English.
Manage paid servers in Billing on desktop and web. Cancel renewal and keep the data, stop a scheduled deletion with Keep server, or delete a server now or after its paid period.
Remove the shadow from confirmation dialogs.
Settings has a Notifications tab for desktop notifications, the test notification and sounds. General is shorter: Language is first, and the settings are in four groups with shorter text.
Consecutive completed runs of one routine show as one row in the chat, with the number of runs and the time of the first and the newest run. Open the row to see each run. A run that fails, stops or needs attention keeps its own row.
Set the default provider for new local agents in Server settings → Providers. Computer permissions remain in Settings → Computer Use.
Open a joined server faster when OpenBot starts. The window no longer waits for the account's server list, the selected server connects while the window loads, and the hidden connection window loads while the session is made. A server that has no saved key still waits for the server list.
Write the time of each step of a connection to a joined server to the local trace file, for diagnostics. The trace does not name the server.
Connect to a joined server faster when OpenBot starts again. The app keeps the server's remote session between runs, so the start asks only for a new ticket. It opens the Signal connection while it waits for that ticket. The kept session is encrypted, and the app forgets it at sign-out or when it starts with no account signed in.
Add Settings → General → Fast connection to servers, on by default. Turn it off to end each remote session when OpenBot quits and to keep nothing between runs, as before.
Write the time when the Signal connection opens to the local trace file (remote-connect:signal-socket).
When you turn on Settings → General → Fast connection to servers, OpenBot also keeps the remote sessions that are open at that time.
When a remote connect fails before it gets its ticket, OpenBot closes the Signal connection that it opened early.
Moved Publish from the chat header to the agent side panel. Click the agent name to find it.
The search field in the sidebar now opens the global search, the same as Cmd+K (Ctrl+K). It no longer filters the sidebar list.
The agent settings panel is shorter. It shows the profile, the instructions, and the model, Knows, Does and Rules groups. Each group row has an icon. Profile, Instructions, Permissions and Advanced open on their own page; the back button returns to the list.
On the web, the loading crew now covers the chat while a hosted server wakes. When the server is online, the crew jumps out and the chat shows.
Test connection no longer opens a browser. It uses the sign-in that this computer already has, and a server that asks for a sign-in shows a Sign in button.
A failed test now says what kind of failure it was: a sign-in is needed, the server refused, the URL is wrong, the server did not answer, it did not start, or it could not be reached.
Group the models of a custom ACP agent by the provider that serves each one. An agent that lists <provider>/<model> names now shows one section per provider in the model picker, instead of a single undifferentiated list.
The changelog page names the mobile app "OpenBot Mobile", because the app is now also on Android. Next to the TestFlight link, a button opens the Android app on Google Play.
The download section of the home page and the /download page have iPhone and Android tiles, in the same style as the desktop tiles. The links to the requirements pages show iPhone and Android in a separate group after macOS, Windows and Linux.
Run independent CI checks and browser setup steps in parallel, with separate logs for each step.
Remove the large package-cache restore from CI setup.
In a file attachment card, the download and open icons are smaller.
Let long message drafts grow to one-third of the window height before scrolling. Keep the cursor visible when the input size changes.
Show install steps for Android in Settings > Mobile Connect. Choose iPhone or Android to see the QR code and the steps for that phone: the Google Play listing for Android, or the TestFlight beta for iPhone. The two cards are numbered, so you install the app first and then sign in. The list of connected phones shows the logo of each platform.
Show the platform logo, a link and the OpenBot logo in the middle of the install QR codes.
Show what a hosted server gives at the top of the add-server dialog: it runs OpenBot 24/7 on an external server, keeps agents and routines running when the computer is off, and needs no technical setup.
Telegram: code blocks in answers keep their language, such as python, so Telegram colors the code.
First run now asks one question, with two equal answers: "Start free", or "I have a subscription" (ChatGPT, Claude or Grok). "Start free" needs no subscription and no sign-in: your agents use the free models of OpenCode, which starts to download when first run opens. Before, you had to find OpenCode at the end of the provider list.
First run has fewer steps. With "Start free", the only other step on macOS asks for the Computer Use permissions. The sound and example-job steps are gone. You can still choose sounds in Settings.
First run changes steps calmly: the old step blurs and fades out, and the next one comes into focus. When setup is done, the setup screen blurs away into the app. With reduced motion, the steps change at once.
When you open OpenBot from first run before the free models are ready, a splash shows: the agents from the first screen jump in a slow wave, and a line under them fills with the download. The app then opens from the splash. Before, a small spinner showed in a button.
The agents on the first screen and on the splash follow the pointer with their eyes. An agent grows a little under the pointer, and nods when you click it. With reduced motion, they stay still.
In the setup provider list, ChatGPT, Claude and Grok show their version under the name, not "Included with OpenBot". Before, an "Update available" badge cut that line off.
The sign-in screen says that the email address also creates an account, and that your chats and files stay on the computer that runs OpenBot.

Removed

Removed the repeated Providers & permissions screen from the account menu. First-run setup remains available.
Remove the outdated Product Hunt launch banner from the landing page.

Fixed

Open Settings on the Hosted servers tab at once from "Manage servers". Before, the dialog showed another tab first and then jumped when the server list loaded. A slow list now shows a loading indicator, and a list that fails to load shows its error on the tab.
OpenCode models with no reasoning setting, such as Big Pickle, Kimi, MiMo and Nemotron, now show "Set by OpenCode" in agent settings. Before, the Reasoning control showed "Medium" as the only level, but OpenBot sent no level and OpenCode used the model's own default. Other ACP agents, such as Cursor, show the same "Set by" text for a model that has no reasoning setting.
OpenCode MiniMax M3 now offers Low (thinking off) and High (thinking on). Before, it offered only Low, so thinking could not be turned on.
Remote hosts can check for compatible OpenBot releases even when they cannot update themselves. Server Settings shows the installed and latest versions, explains the correct update path for hosted servers, system services, containers, and Host Manager installations, and hides installation controls that cannot work.
Server update status refreshes after a lost connection and does not show a reply from a previously selected host.
Show the latest message of a chat or channel in the sidebar as plain text. Before, the preview showed raw Markdown, such as **2 new emails** with its asterisks, while the chat itself showed the same text in bold.
Show "Connecting…" in the sidebar and the message box while a joined server connects after you open OpenBot. Before, the app showed "Create your first agent" and asked you to complete agent CLI setup for some seconds, until the server sent its agents.
Grok chats keep their message order after a reload and when you load older messages. Before, a message that you sent while Grok worked could show above Grok's answer to the message before it.
A message that you send to a running turn now shows after what the agent did before it. Before, it moved up to the turn's first message.
Grok's thinking after a tool call shows as its own step. Before, it joined the thinking from before the tool call.
When you read a chat through an event that came while the agent worked, the agent's answer above it is read too. Before, it stayed unread.
Agents on a custom ACP provider can send messages to teammates and stop their work again. Before, the message failed with "The idempotency key is too long."
Cancelling a browser task no longer shows the raw Codex tool cancellation error. Pending browser work stops, incomplete new tabs close, and existing tabs remain available for the next request.
Reduce idle animation work for active tasks and pending replies. Pause their repeated animations when the window is hidden or the rows are closed.
File chips can now find files outside the workspace from recent ACP and Claude file tool calls. Previously, a chip with only a file name could report a missing file. This lookup is local to the desktop and requires Full access. The last 200 paths per thread stay in memory until OpenBot closes.
Return keyboard focus to the server control after a confirmation closes.
Prevent a server connection from closing with “Signal message is invalid” when the browser sends an empty ICE completion marker.
The web app now opens the last server you selected in that browser. Each account keeps its own selection.
Keep a deleted agent deleted when another device still shows it. Before, opening its chat or marking it read on that device created the agent again, with an empty chat.
Test or delete a webhook routine on a remote server without losing the connection. Before, the app read the host's empty answer as unsafe data and stopped every request to that server until you chose Retry.
Show routine times on a 24-hour clock when the date format of the computer uses one, such as French, Japanese or Russian. Before, the schedule summaries, the time chip and the time editor always used AM and PM.
A server that wakes from sleep no longer shows "Set up" with the provider choice before its agents load. Before, the server sent an empty agent list while it started, until a reconnect sent the agents.
Keep a deleted agent deleted when another device reacts to its message, stops its turn, or steers a queued message into its turn. Before, these actions created the agent again.
OpenBot starts when an unsent message with a skill or MCP server tag is saved in the message box. Before, the app stayed on the startup logo at each launch until you removed the saved draft.
Copy, cut, paste and select all now work in the browser of a remote server, with Ctrl or Cmd and C, X, V and A. Before, the shortcuts did nothing: the page used the clipboard of the server, not your clipboard. The server must also have this update; an older server shows a message that tells you to update it.
Keep the conversation usable when Cursor reports a missing provider session as "Invalid params". OpenBot now uses its saved conversation to continue in a new provider session.
When you close the global search with Escape, focus goes back to the control that had it before.
On the web, a hosted server that sleeps or wakes no longer shows a connection error.
A custom MCP server that asks for an OAuth sign-in no longer shows "The server answered 401. Check the API key or other credentials." It now asks you to sign in. An http:// address for such a server tells you the https:// address to use, and a test from another computer tells you to sign in on the host.
A STDIO server that stops before it answers now says so, instead of a timeout. A command that runs the mcp-remote bridge points you to Streamable HTTP with the same URL.
When Approve, Decline, an answer, Take over or Open details fails in the Dynamic Island, the island now stays open and shows "That did not work. Try again." next to the buttons. Before, the click did nothing that you could see, and the agent stayed blocked.
When a message or file search fails, the search window now says "Search did not finish. Some results can be missing." and shows a Retry button. Before, a failed search showed "No results", and a failed next page stopped the list with no message.
In a language other than English, the Cancel button of many confirmation dialogs, the Copy and Copied labels, and the name that an agent gets when you clear its name now use your language. Before, they stayed in English.
Server settings now ask before they change a member's role or revoke an invitation. A role change can disconnect the member from the server. Before, one click in the menu or on Revoke did the action at once.
Disconnecting 1Password, or an app from the Marketplace, now asks for confirmation first. Before, one click forgot the 1Password token or removed the app and its skills.
The Computer Use rim no longer shows over the OpenBot window. When the window that an agent works in is behind OpenBot, OpenBot now covers that part of the rim, as other windows do. When OpenBot covers all of that window, no rim shows.
Center the agent avatar with the name in the profile card of the agent settings panel, and center the avatar on the profile page.
Antigravity no longer shows "Provider error" notifications for a failed MCP tool call or a timed-out OpenBot browser call. The agent gets the failure as the tool result and can try again.
When an agent calls a memory tool with missing text, text that is too long, or a memory ID that it cannot use, the agent now gets a failed tool result that it can correct. The user no longer sees a "Provider error" notification.
Keep the connection to a server when a desktop client that is connected to it reconnects to the signal server while the connection still works. Before, each such reconnect restarted the connection path, and after 10 restarts the server closed the connection. The chat then stopped for a few seconds while the client connected again.
Let a desktop client stop an agent on a remote server after the computer wakes from a long sleep. Before, the server could close the connection during the sleep while the client still showed it as connected. Each stop then failed with "The WebRTC channel is not open." Now the client connects again when the server closes the connection. When a request could not go out on the closed connection, the client sends it again on the new connection.
An OpenCode CLI update no longer stops with only "fetch failed" after the download completes. OpenBot now gets the OpenCode license from npm, the same place as the CLI, so the update needs no GitHub connection.
When a runtime download request fails, the error now shows the address and the network reason.
A Retry after a failed runtime install now uses the download it already has when that download passes its check, and does not download it again.
Read the history of a custom ACP agent from its process. Before, startup could stop with "Cannot read properties of undefined (reading 'options')" and the local agent backend did not start, and a read of the agent's conversation history could fail with the same error.
OpenBot reads your joined servers again when it opens before your account finishes loading. Before, the first read could stop with "Sign in to OpenBot first.", and the server list did not update until a later check. In 0.33.0, OpenBot could fail to start with this error.
The installed OpenBot web app now opens the app at /app. Before, it opened the landing page.
Connect Figma, or another MCP server on this computer, no longer stops with only "fetch failed". When nothing answers at the local address, the dialog tells you to turn the server on in its app and try again. A connection that this computer blocks, or an address that answers but is not an MCP server over Streamable HTTP, now has its own message.
In an attachment card, the download and open icons now have the same size and alignment.
OpenBot stays open when a Browser tab opens the Chrome Web Store. Before, the app could close a few seconds after the page loaded, and again at each start while the tab was saved.
Keep the message cursor in place when you add or end lists, insert blank lines, or update attachments. Line breaks and multiline paste now work with undo and redo.
Keep text selection and arrow-key movement stable when you use inline tags or leave a suggestion list.
Dragging in the live browser view of a remote member now selects text and moves sliders. Before, the page released the mouse button at the first movement.
When a phone or another computer answers an approval or a question that is no longer waiting, the server now says that the request is closed. Before, it showed a general "Request failed" error.
Show clear loading and recovery states for remote servers on desktop and web. Keep loaded conversations and drafts visible during temporary connection loss.
Retry web connections automatically, report hosted-server start failures, and let Retry start a sleeping server again.
Limit initial remote reads so a server that does not respond cannot leave the workspace loading indefinitely.
Copying selected message text keeps agent, skill, and file references on the same line. Reference names are no longer missing from the copied text.
Show channels in the sidebar again after the desktop app starts, or after it connects to a remote server again. Before, the sidebar showed only agents, and New agent or channel offered only a new agent, until the app window lost focus and got it back.
A phone or another peer that cannot see an agent (a custom ACP agent before protocol 5, or a Cursor or Cline agent before protocol 6) now gets the channel list when one of the channels is led by that agent. Before, the whole list failed and the phone showed no channels. The channel shows with the members the peer can see and no lead.
Saving such a channel from that peer, for example a rename, keeps the hidden members and the hidden lead. A lead the peer picks itself still wins.
Keep remote routine runs quiet through a new optional capability. Older clients still receive the released turn completion event.
Remove Telegram chat links and unused link codes when you remove a server from your account. Other servers keep their links.
An agent with a full memory now gets an error when it tries to save one more memory, and it can merge or forget a memory in the same turn. Before, the new memory was lost when the turn ended.
Show a Gemini rate limit or spent quota as an error that says to wait or choose another model. Before, the turn failed with no message, because OpenBot waited for a usage reading that Gemini does not send. A model that Gemini cannot use and a Google service failure also get their own message.
Say "Not reported" for Gemini, Cursor, Cline and custom ACP agents in the usage menu. These providers have no usage reading, so the menu showed "Unavailable" as if the reading failed.
Stop starting Gemini, Cursor, Cline and custom ACP agents only to read their usage. Each usage check started the provider for a reading that is always empty.
Retry a provider start when model discovery times out, as other start timeouts do.

Security

Update remote-desktop dependencies to fix reported security vulnerabilities.

Oct 7, 2026

Yaji 0.33.0

New

Add a webhook trigger to teammate and group routines. In When to run, open Change trigger and select Webhook. The routine gets its own endpoint and signing secret. A signed request to the endpoint starts the routine. Select Filter events to limit the routine to one event type and to data filters.
The Change trigger menu shows each schedule and the webhook with one line that tells when the routine runs.
You see the signing secret of a webhook routine one time when you save it, and you can make a new secret. When you change the routine to a schedule or delete it, its endpoint stops working.
The routine History shows runs and the webhook requests that the routine ignored. Event data stays on the host computer. The host must be online to receive webhook requests.

Improved

New Claude agents use Haiku 5.5 by default.

Fixed

Show a yellow update notice with an Update button when the selected provider needs a newer runtime. Keep the draft and block sending until the required update is complete, instead of asking the user to reinstall OpenBot.
Hide the failed-send error while the required update notice is shown. Keep Edit and Dismiss available, and restore Retry after the update. Show manual update instructions when OpenBot cannot offer an update.
Pause webhook requests when you sign out. Sign in again to resume the saved routes without changing their URLs or secrets.
Protect saved MCP sign-in credentials from a changed authorization server. Older credentials refresh only when their original authorization server is known.

Oct 7, 2026

Yaji 0.32.0

New

Docker version tags now also accept the v prefix used by GitHub releases.
Report safe provider failure causes and error/warning notifications to OpenPanel, including invalid upload requests from ChatGPT and OpenCode.
Keep bounded local error queues and retry delivery after network failures or restarts. Tracking changes clear pending reports.
Add a separate analytics setting for the browser app. Error reports exclude messages, prompts, paths, and raw exception text.

Fixed

Fixed Codex Computer Use approval storage, including set_value on Windows. Saved tool approvals now remain available after a restart. Configuration failures show recovery steps.
Keep remote Browser Live View available after the host's Team API restarts. Show stream and connection failures instead of a normal end message, and redact secrets from browser start errors.
Explain OpenCode's invalid upload error as a model provider failure, with steps to change models and check saved routines before creating them again.
Docker releases now check that image tags are readable without a GHCR sign-in. Previously, a release could pass while the package was private.
Increase the provider message size limit from 128 MB to 256 MB to allow larger individual records.
Allow model and provider changes after recovery from a stopped provider. Before, a stale active turn could block the change after Codex exceeded the message size limit.
Read provider history in pages and keep completed history on disk. Long chats no longer require a retained copy of the full provider transcript.
Preserve saved messages and attachments when a history import fails or resumes after a restart.
Require Codex CLI 0.156.0 or newer for bounded history reads. The bundled runtime is updated to 0.160.1.
Bundle the internal error-reporting package so the desktop can start correctly.

Oct 7, 2026

Yaji 0.31.0

New

Visual replies. An agent can show a chart, a table, a diagram or a mockup as an HTML page above its reply. The page runs its scripts in a sandbox and can load files from the network. A link in the page opens in your browser after you click it. Agents can look at a page before they show it.
An HTML file opens as a page in the file preview. Use "Show HTML source" to see its code.
Copy a table from a reply as Markdown or CSV.
Click an image in a reply to open it in the image viewer, with the other images of the message.
Play attached audio and video directly in chat, without opening the preview panel.
Cursor and Cline agents on a joined server now show in the app and the browser client, and you can create them there. Before, they stayed on the host computer only.
An owner or admin can download Cursor and Cline on a joined server and sign them in from their own device. Cursor shows a sign-in page that signs the server in by itself, so a server with no screen can use your Cursor plan. Cline shows a device code.
Agents can show HTML pages and Mermaid diagrams in a reply. A code block marked html shows as a rendered page, and a code block marked mermaid shows as a diagram. Use the Preview and Code switch to see the source. Select the expand button to open the page or diagram in a larger window, which grows out of the block. A page runs no scripts and loads nothing from the network. Its links open in your browser.
A chat with three or more days of history shows a day rail on its right edge while you scroll. Each segment is one day, so you can see how long the chat is and where you are in it, also before older messages load: a top segment stands for them. Point at the rail to see the days, and click a day to scroll smoothly to it; the day count rolls to each new number. This is in agent chats, team channels and direct messages on this computer; on a joined server the rail shows the loaded days.
Settings > General has a new "Steer agents while they work" switch. When it is on, a message sent to a busy ChatGPT or Claude agent joins the current work at the next step, and does not wait in the queue. The switch is off by default, so messages queue as before. It also applies to messages from teammates on a server that this computer runs.
Each agent has a "While working" setting in its settings panel. It uses the app default, or it always queues or always steers for that agent.
A message that cannot steer waits in the queue with a "Not steered" label and the reason. Grok, OpenCode, Gemini, Cursor, Cline and custom agents cannot steer a running turn. The message starts when the current turn ends.
A message that could not be sent stays in the chat with Retry, Edit and Dismiss. Retry does not send the message two times, also when the first attempt reached the server and only its answer was lost. A server on an older version cannot detect a repeated message, so for it the chat offers only Edit and Dismiss. If you quit or restart before such a message is sent, its text is back in the composer when you return; nothing sends it again on its own.
openbot.run has a new /providers section with one page for each coding agent that OpenBot runs: Claude Code, Codex, Gemini, Grok, Cursor, OpenCode, Cline, and local models from Ollama or LM Studio. Each page shows how to set the agent up in OpenBot, what OpenBot adds to it, the vendor's own apps and where they run, and answers to common questions. The footer, the provider tile on the home page and the matchup comparisons link to these pages.
openbot.run has a Markdown copy of each guide, news post, comparison and provider page, at the page URL with .md added. /llms-full.txt holds all of them in one file for AI assistants.
The openbot.run FAQ now says what the source license permits.
A folder chip, such as research/eyeliner/, now opens a folder view in the file panel. The view shows the files with their size and date. Select a file to open it, and select Back to go back to the folder.
The file panel can show the source of a Markdown file. It also has Download file for each file, and Show file in Finder on the desktop.
Add a calendar feed for routines in Server Settings > Routines. Apple Calendar (On My Mac) and other calendar apps that read the feed on this computer can subscribe to the next 30 days of runs of active routines, for all agents or for one agent. The private URL works only on this computer, and New URL stops the old one.
Connect a Discord server in Server settings → Connectors → Discord. People mention @OpenBot in a channel, and the Discord Orchestrator agent asks the right agent and answers in the reply chain. Replies to OpenBot continue the conversation. Approvals, Stop, reactions and files work as in Slack.
Connect Bitwarden in Marketplace or Server settings to fill browser passwords and authenticator codes. Share logins through a folder named Shared with OpenBot. The CLI session lasts until disconnect, eight idle hours, or app exit.

Improved

A Mermaid diagram that cannot be drawn shows the parser message. When the diagram module did not load, a retry button shows.
The agent cursor in Computer Use now travels to each point with a smooth, curved move, as a person moves a mouse. A long move across the screen takes a wider arc. Before, the cursor slid in a straight line in a fixed short time. With reduced motion on, the cursor still jumps.
Computer Use now uses cua-driver 0.34.0.
"Use a skill" in the composer's add menu now opens the skill list. Before, it was always unavailable, and you had to know to type $.
The skill list shows skills whose name matches your text first, then skills that match only by description. When two skills have the same name, each row also shows its folder name.
A long reply that streams does less work for each new part. Only the end of the text is read again, the code in a code block is colored one new line at a time, and lines that did not change are not drawn again. The line that is still arriving has no color until it is complete. Before, each new part read and colored the full reply again.
A message you send shows in the chat at once, marked as sending, and the composer stays free for the next one. Before, the composer waited until the server accepted the message. Several messages sent quickly reach the agent in the order you sent them.
In the web client, a send that fails no longer blocks the composer until you check the conversation. The message stays in the chat with its failure.
Use less CPU and memory while channel agents write replies. Before, OpenBot read the whole conversation and the whole channel history again about ten times each second for each agent.
Close the Computer Use highlight windows one minute after no agent controls the computer. Before, they used memory for as long as Computer Use was on.

Removed

Remove the Download all action from chat attachments. Individual file downloads remain available.
The website no longer opens the Product Hunt launch dialog when a page loads. The Product Hunt link above the landing page title stays.

Fixed

Failed Antigravity MCP tool calls no longer show a separate provider error notification.
Remove claims from the website and documentation that OpenBot does not need an account.
Explain an OpenCode model whose provider rejects its API key, such as a Google Gemma model, and tell how to fix the key. Before, OpenBot showed only "Internal error: API key not valid".
The Add member menu in channel settings has a search field, and its list scrolls. Before, the menu had no search and did not scroll, so agents above or below the window edge could not be added.
Sign in and connect to teams on a company network that inspects TLS, such as Fortinet. OpenBot now trusts the root certificates of the operating system, as a browser does.
Show which host a company firewall or proxy blocked when sign-in cannot reach the account service. Before, OpenBot asked you to check that the API was running, and a proxy's block page could sign you out.
Move the window by its title bar when the browser fills the window. Before, only the area under the window buttons moved it. You can also move the window by the header of the browser sidebar.
Long text in chat tables now wraps inside its cell. Before, a table could become wider than the message, and you had to scroll sideways to read it. Long links and code also break inside the cell.
Put the cursor back in the message box when you come back to the OpenBot window. Before, you had to click the message box again before you could type.
A queued message that you steered just as the agent finished its turn no longer shows as done when the agent never read it. It stays in the queue and starts next.
The request that an agent sends to another agent, and the result that comes back, no longer add to the unread count of either agent. Before, an agent showed new messages in the sidebar when its chat had no new message for you.
The skill list now says when the agent has no skills, when no skill matches, or when the skills did not load. Before, the list did not open.
Keep the message you did not send when you go to a different chat. Each agent chat and each channel keeps its own draft. The draft text also stays after you restart OpenBot, for example to install an update. Attached files stay only until you close OpenBot.
The Mobile tab of the changelog shows a new version only when you can install it from the store. Before, it showed the version while the store still reviewed the build.
Attach image now shows PDFs, documents, and all other supported files. Before, on Windows, the file picker showed only images until you selected All files.
A teammate's reply could show in an agent's chat as your own message, with the full teammate prompt. It now shows as a message from the teammate, also for replies that earlier versions saved this way.
When the connection of an ACP agent to its model dropped and the agent tried again, its "API Error" line was joined to the retried answer in one chat bubble. The line now shows as a separate muted activity line.
When a file or folder does not exist, the error now names the path and the agent workspace. Before, a folder chip showed "Workspace path is not a file."
Type the phrase that an input method picks with Shift and a digit, such as an OpenVanilla associated phrase. Before, the message box typed the punctuation of the key, such as "!".
A host that clients reconnect to many times no longer runs out of network sockets and becomes unreachable until a restart. After 10 ICE restarts, a remote connection is replaced with a new one.
Connect again to a joined server after the computer wakes from sleep. Before, the server could stay unavailable until you restarted OpenBot.
Sign in to Framer in the embedded browser. Before, Framer refused every sign-in with "Cannot log you in" or "Verification failed", because the browser identity included the OpenBot name.

Security

The Windows installer and app are now signed by SYNTHETIFY LABS SPÓŁKA Z OGRANICZONĄ ODPOWIEDZIALNOŚCIĄ. Windows no longer shows "Unknown publisher" for them, and later updates install only when they have the same signature. Before, the Windows release was not signed.
The account service now checks the skills and agent marketplace admin token in constant time. Before, the check time could show how much of a guessed token was correct.
Bitwarden fills require an exact HTTPS origin and a current item in the shared folder. Session keys and login values are not saved by OpenBot, and short authenticator codes are redacted from logs.

Oct 5, 2026

Yaji 0.30.0

New

An agent can suggest a Marketplace app in the chat. The card shows the app, its state, and Connect. Connect opens the app in Marketplace and starts its sign-in there. You can dismiss the card and undo.
When an agent needs GitHub and it is not connected, the agent shows a GitHub card in the chat.
The openbot.run home page shows what OpenBot does, answers common questions, and links to each comparison.
An official Docker image, ghcr.io/nightly-labs/openbot, for linux/amd64 and linux/arm64. It runs OpenBot as a self-hosted server with the Electron sandbox on. Sign it in with docker exec -it openbot openbot login. See docs/docker.md.
openbot.run has a download page for each system: /download/macos, /download/windows and /download/linux. Each page gives the system requirements, the install steps and the known limits, and /download links to all three.
openbot.run has a download for Linux on arm64.
openbot.run has four new pages in /compare: Codex vs Claude Code, Claude Code vs Antigravity, Cursor vs Claude Code, and the best AI agent apps of 2026. Each pair page compares the two apps from their official sources, and shows how OpenBot runs both as one team. The roundup lists 13 apps and links each one to its comparisons. The Codex, Antigravity and roundup pages show a chart of scores, cost per task and time per task from the Artificial Analysis Coding Agent Index.
Connect 1Password in the Marketplace or in Server settings > Connectors. The page shows three steps: Install puts the 1Password CLI in place for you, with no administrator password; then turn on the CLI integration in the 1Password app, which the page sees when you come back; then Connect 1Password creates a "Shared with OpenBot" vault and a service account that can read only that vault. You can paste a service account token instead. Agents then sign in to sites in the OpenBot browser with the logins you move into that vault. The browser fills the password or authenticator code; agents never see it.
Full Turkish language localization across OpenBot, covering 100% of messages including new features, connectors, Marketplace v3, schedule calendar, and all error catalogs.
Choose the message send shortcut in Settings > General > App behavior (desktop) or Account > Preferences > App behavior (web): Enter to send, or ⌘Enter on macOS / Ctrl+Enter on Windows and Linux to send with Enter adding a new line. The choice stays on the device and browser. An Enter that confirms IME text never sends.
When a Claude or Codex account reaches its usage limit, OpenBot shows one desktop notification for the account, with the number of agents that wait and the reset time when the provider gives it. Each agent on that account shows "Waits for limit" and the reset time in the sidebar.
A routine has a new setting, "If the account is at its limit", on this computer: wait and run after the reset, or skip the run. Wait is the default.

Improved

On openbot.run, News, Guides, Compare and Plugins in the header now open their pages when you click them. Hover over them to see the panel.
The openbot.run footer names Synthetify Labs as the owner.
A chat now shows consecutive messages to and from other agents as one row, such as "7 messages with 7 agents". Click the agents to see each message.
The openbot.run footer links to the self-hosted server guide.
The direct installer links are now /download/<os>/latest. An old /download/<os> link opens the download page for that system, not the installer.
The Schedule view of a server is now called Routines, the same name as in agent and channel settings.
A failed task card in the Dynamic Island now has a "Dismiss" button. The button removes the card from the island only. The conversation keeps the error, and the agent list keeps the failed mark. The island then shows the newer replies of the other agents (#1348).

Removed

Removed the public guide-authoring page. Its instructions now live in the contributor documentation.

Fixed

The openbot.run home page title reads "Meet OpenBot" to search engines, and its text shows when JavaScript is off.
The openbot.run home page title shows sooner: it no longer waits for the page script.
www.openbot.run and http://openbot.run now go to https://openbot.run.
Page addresses that end in / now redirect permanently to the address without it.
The guides, news, plugins and changelog pages have clearer titles in search results.
The server member count now uses the singular form for one member. Before, a server with one member read "1 members" in English.
The Dynamic Island toggle now reads "Expand" and "Collapse" in the interface language. Before, it always used these two words in English. A screen reader then read English inside a control that was otherwise in the chosen language.
The composer error banner showed two English messages that no catalog owned. With the interface in another language, they stayed English above a translated composer. Both now come from the composer catalog. The attachment limit message also reads the limit from the shared limit, so a change to that value updates the text.
The desktop app no longer closes when a remote server announces a new name and the saved server list cannot be written, for example when the profile directory is read-only or the disk is full. The new name still appears in the app, and the next write of the list saves it.
A plugin listing with a server argument that is not text now stops the plugin catalog build. Before, the build dropped the whole argument list, so the listing shipped a server that started with no arguments.
The app no longer freezes when an agent declares a very long task resource list. OpenBot now refuses an oversized list at once, instead of resolving every path in the list first.
Signing out, disconnecting or quitting now ends a file download that is still waiting. Before, the download waited for its full 60 s timeout and its timer kept OpenBot from closing.
Closing a remote connection while Signal checks its ticket no longer uses one of the account's remote connection slots. Before, the closed connection kept its slot, and the desktop was told that a phone was ready. A later device was then refused because the account was full.
A remote client whose agent snapshot is too large to send no longer stops the other remote clients from receiving that agent event. Before, every client that connected after it missed the event and kept a stale queue until the next event arrived.
A chat could stop updating on screen when notifying the window failed. Every update now runs, so the chat stays the same as the saved messages.
The skill marketplace no longer skips skills when a client pages with a cursor from before cursor v1. That cursor held only a timestamp. The catalog order also leads with the featured flag, so the timestamp could not say where the last page stopped. The page now comes again and the rest of the catalog follows.
The agent marketplace no longer skips agents when a client pages with a cursor from before cursor v1. That cursor held only a timestamp. The catalog order also leads with the featured flag, so the timestamp could not say where the last page stopped. The page now comes again and the rest of the catalog follows.
On macOS, the message that blocks an update while another OpenBot session runs now tells you when the other session runs in your own user account. Before, it always told you to stop OpenBot in every other macOS user account, even when the blocking process ran in your own account.
Open Desktop now shows why a remote desktop did not start: Sunshine did not start, Moonlight Web did not start, or pairing failed. Before, it showed only "request failed".
When Sunshine stops on the host, OpenBot ends its remote desktop sessions and starts Sunshine again for the next session. Before, each new session stayed at "Connecting".
Open Desktop shows an error when the host does not start the stream in 60 seconds. Before, it stayed at "Connecting" until you closed it.
Signal no longer refuses a new phone or browser with "Too many active remote connections." when one account has several computers and devices. The account limit is now 32 Signal connections, not 8. A phone that reconnects after a network change no longer loses its own place to its old connection.
After a restart, a Claude agent chat no longer shows the context summary, "[Request interrupted by user]", or the output of a command such as /compact as a message from you, and no longer repeats an answer it already showed. The next restart also removes these messages from chats that already show them. An answer that was already repeated stays.
You can install the OpenBot website as an app from Chrome, Edge and Android. Before, the site manifest told the browser to keep the site in a normal tab, so the browser did not offer to install it.
On a phone, the email sign-in code now fills from the keyboard suggestion or from autofill, also with its hyphen. Each character typed on an Android keyboard now goes into the next box. Before, each character replaced the first box.
Canva opens in the embedded browser. Before, Canva showed an "Update your browser" page and did not let you edit a presentation, because its server refused the OpenBot token in the user agent.
Gemini agents no longer fail every turn with "Invalid value at 'tools[0].function_declarations[…]…enum[0]' (TYPE_STRING)". Two OpenBot tools sent numbers in a list of allowed values, which Gemini accepts only as text.
When you change an agent's provider, the new provider now gets the work the previous one did: the commands it ran with their results, the files it changed, and its progress notes. Before, it got only the chat messages, so it did not know much of the earlier work.
When you ask an agent on a connected server to create a routine, the routine now runs in your timezone. Before, it ran in the server's timezone, so a routine for 8:00 could run hours early or late. This works from the desktop app and the web client when both the app and the server are updated.
When you change an agent's provider, OpenBot now saves the earlier provider's work log at the moment of the switch. Before, if that provider stopped soon after the switch, as an unused provider does after a minute, the new provider could get the chat without the earlier work.
A custom agent that serves one folder for each process, such as Command Code (cmd acp), lists its models while a bot uses it, and runs bots in more than one folder. Before, the model list failed with "This cmd acp process serves ...; start another for ...", because OpenBot used one process of the agent for all folders.
A steered message to a custom ACP agent that accepts one prompt at a time is not lost. Before, the agent refused it with "A prompt is already running for this session", OpenBot showed a provider error, and the message was not sent. Now OpenBot sends the message when the running reply ends.
On the website, the page keeps its scroll bar and does not move when you hover over or focus "App" in the header. Before, the scroll bar went away, the page moved sideways and stopped scrolling.
Codex and ACP agents now receive a clear reason and can retry when an agent profile request is invalid, for example when profile instructions exceed the length limit or a model is not available, instead of receiving a generic provider error.
The files of an "Update from" message now sit at the left edge, under the update, where an agent's own files sit. Before, they were pushed to the right.
A chat row no longer takes the data of the next message when messages in the middle of the chat change order. Before, a row could hold the raw text that an agent sent, or the file of another message, until the list updated.
A chat message that cannot be shown now shows "This message could not be shown". The rest of the chat stays usable.
A message, a routine run or a local script run that reached an account at its usage limit failed with "Internal error: You've hit your session limit", and OpenBot did not run it again. Now it waits in the queue and starts after the reset. A turn that had already run a command still fails, so the command does not run twice.

Oct 2, 2026

Yaji 0.29.0

New

Compare OpenBot with ChatGPT dots on openbot.run/compare.
Each comparison on openbot.run/compare has a "Where you can use it" row. It shows the countries and languages of each product, and that OpenBot has no region lock.
A script on your computer can now run an agent's routine and give it a payload, for example to wake the agent when a long build ends. Turn on Local scripts in the agent's settings, then use Copy run command on a routine. The setting is off by default.
The openbot.run header menu has a Compare section, and each comparison links to the other comparisons.
Install OpenBot as a server on a Linux computer with no screen, such as a VPS, from a terminal: curl -fsSL https://raw.githubusercontent.com/nightly-labs/openbot/main/scripts/install-server.sh | sudo bash. Sign it in with sudo openbot login, and use it from the desktop app, the iPhone app or openbot.run/app. The openbot command also shows the status and the log, changes the server name, installs updates, and removes OpenBot with sudo openbot uninstall. The install also works on Debian 12, Debian 13 and Ubuntu 26.04. See docs/self-hosted-server.md.
OpenBot writes provider diagnostics to logs/providers/providers.log in its data folder: state changes, CLI checks, start times, model list results and provider errors. OpenBot removes credentials from each line, and does not write environment values, tool output or conversation text to the file.
The provider settings show the last error of each provider until the provider lists its models again, with a "Copy diagnostics" action.

Improved

The comparisons on openbot.run/compare include Cursor CLI and hosted OpenBot servers in the EU, Claude Cowork's move to cloud-only tasks on 6 October 2026, and Manus 2.0.
News, guide and comparison pages on openbot.run load faster: their artwork is WebP and about 3% of its old size.
Search results show fuller descriptions of the plugin pages and shorter descriptions of the comparison pages on openbot.run.
"The selected agent model is unavailable." now gives the cause: the provider is not connected, the provider listed no models (with its last error), or the provider does not list that model.

Oct 2, 2026

Yaji 0.28.0

New

Your agents can answer in Slack. Open Server settings → Connectors → Slack, select Connect Slack one time and install OpenBot in your workspace. People mention @OpenBot in a channel or send it a direct message. OpenBot joins every public channel by itself; a private channel needs /invite @OpenBot. The connect dialog also adds the Slack Orchestrator, an agent on the model you pick, in a collapsed Integrations section of the sidebar: it receives each request, gives the work to the agent that fits best, and posts the answer in the thread. The answer comes in the same thread, can include files, and asks the person who wrote for approval with buttons. Reply stop to stop a request. Slack's events reach this computer through OpenBot's Signal service, which checks them and passes them on without storing them, so this needs an OpenBot account and a name for this computer. A hosted server stays awake while Slack is connected.
Brazilian Portuguese for the desktop app and web client. Select Português (Brasil) in Settings, or use your system language.
More of the desktop app and the web client is in Turkish, including server settings, providers, marketplace, routines, usage and remote desktop.
Cline is a provider. Its free models have a limit for each model, and the provider of a free model can use your prompts to train models. Select Download on the Cline row in More providers, then sign in with your Cline account in the browser, or set CLINE_API_KEY. Cline agents stay on this computer: team members do not see them.
You can run your own account service, Signal and TURN for remote access. Start the app with OPENBOT_AUTH_API_URL and OPENBOT_REMOTE_SIGNAL_URL. Invitations for your service are openbot://join links. See Self-hosted remote access. (#1263)
A Schedule view shows the routines of all agents and channels on a server in a day or week calendar, with past results and planned runs. Open it from the server menu or the command palette.
Restart a provider from its menu in Settings, and restart all custom agents from the Custom agents section. OpenBot waits until no agent of that provider works, then starts the provider again and reads its version, sign-in state and models again. Agents of other providers continue to work, and messages sent during the wait run after the restart.
Restart OpenBot when no agent works, from Settings > Updates. New routine runs wait until the restart and run after it. When an update is downloaded, the same action installs it.
Add sound feedback. When you turn it on, short sounds play across the app: for buttons, switches, tabs and menu choices, for dialogs and menus that open and close, for moves between agents, channels and servers, and for typing. Other sounds tell you that an action worked or failed, that an agent waits for your approval or answer, that a run failed, and that an attachment import finished. Choose Warm, Mechanical, Bubbly or Clicky sounds, or Off, in a new step of the first setup, or in Settings > Notifications on the desktop app and the web app. Click a choice to hear a preview again. It is off by default.

Improved

Server settings → Connectors lists GitHub and Slack. Select one to open its page.
The logo and the greeting blur a little while the Dynamic Island changes size, as much as the move is large: one slow step of the width setting barely blurs them, and a large move blurs them in full. The logo also gets a little smaller and half closes its eyes, and on a large move rounds its corners a little. The greeting gets a little fainter, and when a large move lands it pops softly. Both are sharp and at rest again as the island settles. With Reduce motion on, none of this happens.
Each idle greeting is now a small 3D card that moves once as it shows, and then turns back to itself: the hand waves, turns into heart hands and back; the smile spins like a ball into a grin and back; the raised hands turn to clap twice and turn back raised; and the sparkles light up one star at a time and then shimmer. One greeting hands over to the next with a soft blur, fade and scale. The greeting is a little larger (16px). With Reduce motion on, the greetings do not move.
Below 100%, the idle Dynamic Island width now changes with each step of the width setting, from the smallest island at 20% to the default at 100%. Before, the lowest steps gave the same island. On a built-in display with no notch, the smallest idle island is now 84px, with a 16px gap between the logo and the greeting. It was 120px. The same percent can give a different width than before. Beside a physical notch, the width does not change.
The notice about different OpenBot versions on a server now says which side to update: OpenBot on the server, or this app. It offers the host update only when the server is the older side.
A hosted server now stops about 15 to 30 minutes after your last message, change or typing, also when an app is open. Before, an open app that nobody used kept the server running for 1 hour or more. While the server is asleep, you can still look at it: a moon beside the server name shows that it sleeps. Press a key or click to wake it. Your draft stays.

Fixed

A custom ACP agent that is slow to start now shows its models. Model discovery waited only 5 seconds for it. (#1230)
The Open OpenBot button on the page that Slack returns to after you install OpenBot is now in the middle of the card.
A link to a file that an agent edited now opens the file. Before, a link to a file outside the agent's workspace, a link with a line number such as page.tsx:12, a ~/ path or a file:// link showed "not found" or an error. A file outside the workspace opens only for an agent with Full access, and "Open file externally" shows it in the file manager. (#1240)
A link to a file that was moved or deleted now says so. Before, it told you to ask the agent to create the file.
A long error after a failed provider update ran out of its notification. Now it stays inside, stops after three lines, and "Show details" shows the full text.
When the Dynamic Island width or height changes, the logo and the greeting now move with the black island, on the same curve and at the same distance from each edge. Before, they got to their new place first and showed outside the island while it grew.
Settings → Dynamic Island → Size: the built-in display preview now draws the island that display shows. On a built-in display with no notch, it no longer draws a notch. Each preview scales so that the widest width fits its frame, and keeps that scale while the setting changes. The preview island hangs below the frame edge, so the edge runs over it.
A lower Dynamic Island now has smaller bottom corners (7.9px at 75% height, 14px at 100%), and a higher one larger corners. On an external display the island stays a capsule at every height. At 100% height nothing changes.
With the Cursor provider connected, the model picker showed no models for any provider. Now the Cursor models load and you can select them. A model with an id that OpenBot cannot use no longer removes the models of all providers.
When a custom ACP agent is slow to list its models, OpenBot now closes the session that it opened to read the list. Before, each slow model list kept one idle agent process open until OpenBot quit.
When OpenBot stops an ACP agent (because it is idle, to restart it, or at quit), the errors that the agent writes while it stops go to the log. Before, OpenBot showed each of them as a "Provider error" message. When an agent stops on its own, OpenBot still shows its errors.
A provider CLI download no longer fails with "another instance is replacing it" when no other OpenBot runs. When another program holds the new files open, OpenBot now waits for a moment, and then tells you to close that program. (#1264)
A failed update notification now closes when the CLI is updated later. Retry on a provider row now removes the old error text, and the custom provider row no longer shows a second Retry next to Add. (#1264)
On a busy computer, a provider CLI that answers slowly no longer shows as broken. OpenBot now waits 10 seconds for its version, says that it did not answer in time, keeps the last known version, and tries again by itself. (#1258)
A provider refresh no longer marks a connected provider as failed while its models stay in the model list. (#1258)
A desktop chat no longer shows an empty bubble when an agent ends its turn without an answer, for example after it reads a teammate update that needs no reply. The phone app and channels already hid these rows.
Keep OpenCode's free model picker available when a temporary model-discovery failure occurs.
On Windows, an agent that starts from a .cmd or .bat file now gets each argument as written. Before, the command processor split or changed an argument with a space, a quote, & or %.
An agent request that failed no longer shows an Error handling request { toast, and the lines under it no longer show as more toasts. The failure shows in the chat with the reason that the agent gave, and the full text goes to the log.
Gemini no longer fails with "Download failed" when its first start after the download is slow. OpenBot now waits up to 3 minutes for a CLI that it just installed to start, and onboarding shows why a download or that first start failed.
When more than one provider updated at the same time, the update notifications jumped and overlapped. Now they stay in place while the downloads continue.
On Windows, a provider CLI install (for example OpenCode or Gemini) no longer fails with "another program has its files open" while Windows Defender scans the new files. OpenBot now waits up to 60 seconds for the scan to end.
A custom agent kept its saved model after a restart. Before, a custom agent that started slowly could move its agents to its default model.
The Usage panel shows the remaining Grok usage. Before, a Grok account with no use in the current period showed a dash in place of a percentage.
A provider row in the Usage panel shows "Loading…" while usage loads, and "Unavailable" with "No limit reported" when the provider reports no limit. Before, both showed a dash.
A Claude Code agent no longer shows a <task-notification> block as a message from you after OpenBot restarts. These blocks are notices about background tasks. Notices that earlier versions added are removed the next time OpenBot starts.

Oct 1, 2026

Yaji 0.27.0

New

Turkish (tr) language support across desktop, web, and shared packages.
Cursor is a provider. Select Download on the Cursor row in More providers, then sign in with your Cursor account in the browser, or set CURSOR_API_KEY. OpenBot uses a cursor-agent that you installed yourself until a download exists. Cursor agents stay on this computer: team members do not see them.
Settings → General → Appearance has a Logo color setting with 10 colors. The chosen color shows on the Dock or taskbar icon, on the Dynamic Island and on the logos in the app. When OpenBot is closed, the Dock shows the lavender icon, because macOS has no alternate app icon for desktop apps. A dev or preview build keeps the color of its build.
Global search (⌘K) finds channels, files, routines, commands and settings pages. New filters show Channels, Files and Routines. Select a file to show the message that has it. Select a routine to open its settings. Commands such as New agent and New channel, and each settings page, show in All when the query matches. Files come only from this computer, so a joined server and the web client do not show the Files filter.

Improved

The GitHub page in Server settings > Connectors has a new design. It shows the GitHub status and your account picture. A dialog guides you through the sign-in code. You can filter the list of repositories. Disconnect asks you to confirm first.
When an account has the maximum number of paid hosted servers, the add server dialog now tells the user before they choose a plan. The plans are disabled, and a "Manage servers" button opens the list of hosted servers, where the user can delete one.
Hosted sites now belong to the server that published them. Each server has its own limit of active sites from its plan: 1 with no plan, 3 on Starter, 10 on Standard and 50 on Pro. Uploads from a computer that is not a registered server share one site for each account.
Manage sites in Server settings > Sites, on the desktop and in the browser. All members see the list. Only an owner or admin can delete a site. The Hosted sites tab in Settings is removed.
After a downgrade, sites above the new limit stay until they expire. The server cannot publish a new site until it is below the limit, but it can update an existing site.
Manage AI providers in Server settings > Providers, not in Settings. Each server shows its own providers. For a server that is not active, the section offers to switch to it.
Global search (⌘K) has a new, more compact design. Each result uses one line, and All shows each kind of result in a separate group. The query text is highlighted in the results. The filters are in the search field, and Tab and Shift+Tab change the filter. A footer shows the keyboard shortcuts, and a spinner shows when results are slow to load. The search opens, closes and changes height with a short animation, the results fade at an edge that has more to scroll, and the highlight and the filter pill slide to the new row or filter. Long result lists scroll smoothly with the arrow keys and stay fast with thousands of results. The Messages and Files filters load more results as you scroll to the end.
The Marketplace is now one full-screen window with the tabs Agents, Apps and Skills, a search field and a filter menu. Each agent, app and skill has its own page.
You can install a skill on all your agents in one step. The Marketplace names the agents where the install failed.
When you add an agent, the Marketplace stays open. The button then says "Open chat".
The GitHub page in the Marketplace and Server settings › Connectors show the same GitHub connection.

Removed

The Cursor preset for custom ACP agents. Use the Cursor provider. A custom agent that you saved with the preset continues to work. If its ID is cursor, you cannot edit it: to change it, remove it and add it again with another ID.
The GitHub plugin that used a personal access token. Use the GitHub connector. A server that the plugin added stays, and shows as a custom server. Remove that server, so that your agents use the GitHub connector.
Skill and agent publishing from the desktop Marketplace.

Fixed

Search now finds a phrase that a line break splits in a message, in the chat and in global search. The chat search also highlights it. Thanks to @aniruddhaadak80 for the first fix in #1174.
With Grok, an MCP server from ~/.claude.json, ~/.cursor/mcp.json or .mcp.json that needs an OAuth sign-in no longer shows a "Provider error". Before, Grok's worker quit with fatal … AuthRequired line showed as an error, but the chat worked without that server.
An agent that another agent creates now gets the provider, model, and reasoning effort of that agent, unless the request names different ones. Before, each new teammate started on the default model. (#1201)
Keep the other devices connected to a desktop when one phone reconnects. Before, a late network message from the old connection of that phone could disconnect every device from the desktop.
Keep the web client connected through a short network change. Before, it made a new connection after 5 seconds without an ICE restart, and it retried only every 10 seconds.
The startup screen now shows the full logo animation before it fades. Before, on a fast start, the logo flashed half drawn and was gone.
A cancelled or unfinished Stripe payment no longer blocks a new hosted server. A new plan choice now changes the plan of the server that waits for its first payment, and does not add a second server.
Start the Claude and Codex CLIs on Windows when the user name holds a space, such as C:\Users\Jane Doe. Before, the Claude sign-in check failed with "'C:\Users\Jane' is not recognized as an internal or external command".
Pressing Enter to confirm Japanese, Chinese, or Korean input in the custom answer field of an agent question no longer sends the answer. Before, the answer was sent with the unconfirmed text.
Keep the last successful OpenCode model list when a later catalogue refresh fails. This keeps a connected provider's model picker usable after a temporary ACP timeout (#508).
Notifications that arrive together merge into one stack, and hover shows all of them with no empty space. Up to six update notifications show, not three.

Sep 30, 2026

Yaji 0.26.0

Action needed after you upgrade

A hosted server updates itself. It downloads a new OpenBot release in the background and starts it at its next start, so an open session does not stop. A server that was set up before this release has no updater, so it stays on its current version. Upgrade such a server one time by hand, as `docs/hosted-servers.md` says. The data in its home folder and in /srv stays.

New

In an agent chat on a team, each message from another person stands on the right, with your messages, and shows their name. Their bubble has their own color, so it does not look like yours. Agent messages look the same as before. Messages sent before this update show as yours.

Improved

A routine run in an agent chat shows as one short line with the routine name and its state. The chat no longer shows the routine instruction as your message. Open the routine to read its instruction. Chat search no longer finds routine instructions.
When a hosted server is low on memory, new messages wait in the queue and start when memory is free. The agent shows a notice, and the browser opens no new tab.
A hosted server runs at most 4, 8 or 16 agent turns at the same time, from the memory of its plan. Other messages wait in the queue, and a message from a person starts before routine runs and teammate messages.
A plan change no longer stops a hosted server that is in use. The server moves to the machine of the new plan when it has no use for 7 minutes, or at its next start.
An agent's browser tab that nobody uses for 30 minutes now unloads its page to free memory. The tab stays open with its URL, title and preview, and the page loads again when the agent or you use the tab. On a server that is low on memory, this occurs after 5 minutes.

Fixed

In the browser live view of a remote server, Enter submits a form and breaks a line, and Backspace, Delete, Tab and the arrow keys work.
With OpenCode, a Computer Use window read adds much less text to the conversation, so long Computer Use sessions stay faster and compact less often. Before, each read of a large window added a list of element tokens about three quarters of the size of the window tree.
The Computer Use driver stops when OpenBot stops, also after a crash or an out-of-memory kill. Before, the driver kept running and used memory until you logged out. At the next start, OpenBot also stops the driver that the last crashed run left running.
OpenBot uses less memory in long sessions. It no longer keeps a second copy of each conversation, and it releases the data of each finished turn.
The picture-in-picture browser controls release their memory when the system closes the window.
Antigravity and custom ACP agents stay stopped when they are idle. Before, the usage display started them again every five minutes, and a custom agent router then started the process of each custom agent.
On a new hosted server, OpenBot keeps running when the server is out of memory. Before, the system could stop OpenBot, or stop all of OpenBot when one agent process used too much memory. Now an agent process stops first.
The Dynamic Island shows new messages again when you have more than 10,000 unread replies. Before, it stopped updating, and OpenBot logged an error each time the island changed.
Each Dynamic Island window uses about half the memory. Before, each window loaded the full app. There is one window for each display.
OpenBot keeps at most 16 chats in memory that you only read. Before, each chat that you marked read stayed in memory until you quit, about 1 MB for a long chat.
On a hosted server that is out of memory, the kernel now stops an agent process before the processes that OpenBot uses to open browser tabs.
Sending a message, or a change to a queued message, no longer writes all message history to the database again. Before, each change wrote every message and delivery, so on a server that ran for weeks each change became slower and used more disk.
OpenBot stops Codex, OpenCode, Grok or a custom agent when it sends one message larger than 128 MB, and shows the reason. Before, OpenBot kept all of the message in memory with no limit, and the system could stop OpenBot on a hosted server.
OpenBot uses less memory when it runs for a long time with many agents. When an agent chat is not used for 10 minutes, OpenBot removes its copy from memory and reads it again from the database when you or the agent use the chat. The chat does not change.
A hosted server uses much less memory when a teammate or a remote device downloads or uploads a file. Before, one 100 MB file could use about 200 MB of memory for the full transfer. Now the server sends and receives files from disk. A server receives two attachment uploads at the same time. Other uploads wait for their turn; they do not fail.
With Codex (GPT models), an agent now gets the tools of an MCP server whose name has a space or another character that is not a letter, a digit, _ or -, such as "Home Assistant". Before, the connection test passed but the agent got none of the server's tools.
channel_forget_memory now matches the saved memory text regardless of letter case. Before this fix, forgetting a channel memory with different casing than the saved text (for example asking to forget "use bun for scripts" when the saved text was "Use Bun for scripts") silently failed and left the memory in place.
When the live view of a remote host's page ends or fails, the panel now shows the message in the language you selected. Before, it always showed the message in English.
With automatic downloads on, one restart now installs the newest OpenBot version. Before, OpenBot stopped checking for updates after a download finished. If a newer version shipped before the restart, OpenBot installed the older download and offered one more update after the restart.

Sep 29, 2026

Yaji 0.25.2

New

Every member of a joined server can import agents from Grok Bot, not only the owner of the computer. Open Server settings → Import on a joined server in the desktop app or the browser client. The export must be a .zip under 100 MB. When the server already has a skill of the export, the agent of a member uses that skill and the import says so.
The host keeps the Lock Screen activity of a joined iPhone current while iOS stops the OpenBot app in the background. It encrypts each update for that phone before it goes through the OpenBot account service and Apple, so neither can read it. The host keeps the phone's push token in memory only, for the session that gave it.
Enter a promotion code on the payment page when you start a plan for a hosted server.
Connect an AI provider before you make the first agent on a server that you own or manage. Before, a new server opened on the agent form with no provider connected.
Sign a server in to Claude and Grok from your own device, when the server has no browser that you can see, such as a hosted server. Claude shows a code on its page that you paste into OpenBot.
Connect GitHub in Server settings > Connectors. Every agent on this computer then gets the GitHub tools, and gh and git sign in as you, in the repositories where you install the OpenBot GitHub App. The panel lists these repositories. You do not need a personal access token.
In the repositories where you can push, GitHub shows the issues, pull requests and comments that an agent makes with the GitHub tools as openbotgit[bot], not as you. OpenBot gets short-lived tokens for this from the OpenBot account service, which keeps no token. gh still acts as you.
Use remote desktop on a Linux x64 host in an X11 session, including a hosted server. Wayland is not supported.
On a hosted server, remote desktop shows the full server desktop with its window manager, so a click moves between windows.

Improved

In a dev or preview build, the OpenBot logo on the Dynamic Island has the color of that build, as the Dock icon does. Two OpenBot apps that are open at the same time are easier to tell apart.
Show a new loading screen in the browser app: a small crew of agents hops while OpenBot loads, and jumps out when it is ready.
Show the OpenBot logo while the desktop app starts, and fade it out when the app is ready.
Show how long the agent's activity line has stayed the same when it stays for more than 5 seconds, so a slow step no longer looks like a stopped agent.
Show "Using an app on this computer…" while a Computer Use action runs, and "Deciding the next step in the app…" while the model chooses the next one. Before, both showed a general tool text.
Tell agents to go directly to the named application and action with Computer Use, without listing other applications or reading the same window again.
Log the time that the Computer Use driver takes to answer each call, and each call that gets no answer, so a slow step shows whether the driver or the model used the time.

Fixed

The logo color of a dev or preview build matches its app icon. Before, the logo in the app was orange or bright green, and the app icon was gold or soft green.
A Grok agent answers again after you sign in to Grok with a different account or change the xAI API key. Before, each message in an earlier chat failed with "reasoning encrypted_content was not issued to this caller". Now OpenBot starts a new Grok session that keeps the chat history, and sends your message again.
On Windows at a display scale other than 100%, the Computer Use border goes around the full window. Before, only the top and left edges were on the screen.
The Computer Use border stays on the window while the agent thinks between two steps, and goes away when the turn ends, fails, or is cancelled.
When an OpenCode model request fails, the error tells why: a rate limit, a billing problem with the provider account, a failure on the provider's side, or no network connection. Each error tells you what to do, and whether waiting helps. Before, every cause showed as "Internal error:" followed by the provider's text (#1163).

Sep 29, 2026

Yaji 0.25.1

New

Publish an agent from the browser client. An owner or admin of a server can publish, update or unpublish the share link of an agent on that server from the conversation header. The template belongs to the account signed in on the host, and the host must run this version or later.
Open Account settings from the account menu in the web app. It has your profile, your hosted sites (open or delete them), and preferences for the language and the completion sound.

Fixed

Save the "Play a sound when a task finishes" setting in the desktop app. Before, the sound played when the setting showed off, and the setting did not stay after a restart.

Sep 29, 2026

Yaji 0.25.0

New

Update OpenBot on a joined server from Server settings > Updates. Owners and admins can check for an update, download it, and restart the host when its agents are idle, or restart it at once. The version notice for a server has an "Update host" button.
Turn off updates from server admins in Settings > Updates on the computer that runs the server. That computer shows who asked for an update and can cancel the restart.
Turn on automatic download and automatic install when idle for a joined server in Server settings > Updates.
A notice shows the download percentage while a server updates, and tells admins when a server has a new version.
All members of a server see a notice before and while it restarts into an update. OpenBot connects again when the server is back.
Chat messages show LaTeX math as typeset formulas. Write inline math as $...$ or $...$, and display math as $$...$$, \[...\] or a math code block. A formula that does not parse shows its source. Copying a formula gives its LaTeX source. Before, the formula showed as raw source with its dollar signs and backslashes.
The page of a shared agent has an "open in browser" link below the card, next to "Download it" and "open the app". It adds the agent from the browser client when you do not have the desktop app.
The sidebar shows agents that wait for you in a "Needs you" group at the top, with a count. Each row shows what the agent waits for, "Answer", "Review" or "Take over", and a tooltip shows the question or the command. The agent goes back to its section when you reply.
The message box continues a Markdown list. Press Shift+Enter after a - , * or 1. item to start the next item. Press Shift+Enter on an empty item to end the list.
Your own messages show Markdown, such as lists, bold text and code, as on the iPhone app.
Show browser notifications in the web app when an agent finishes, stops with an error, or needs your input or approval, on every connected server. The web app shows them while its tab is open and does not have focus.
Mute a server and set its notification level in the web app, from the server menu or Server settings. This browser keeps the choice.
Explain how to use the OpenBot Marketplace to install agents, plugins, and skills, and submit an agent for review.
Add "New agent" and "Mark all as read" to the sidebar menu that opens when you right-click an empty part of the sidebar, on the desktop app and in the browser client. "Mark all as read" marks all agent chats and channels as read.
Add a hosted server with the plus button in the server rail. Choose a plan and pay on the Stripe page. OpenBot sets up the server when Stripe confirms the payment, and connects to it. Each hosted server is a Linux machine.
Manage the plans of your servers. Desktop Settings has a Billing tab, and the web app has Billing in the account menu. It shows the plan, the storage, the price and the renewal date of each server that your account pays for, and a warning when a payment failed.
Each server has a menu to change or cancel its plan, or to renew a plan that ends. Payment method and invoices opens the Stripe Customer Portal. You enter card details on the Stripe page, not in OpenBot.
The plan of a server sets its member limit: Starter 3, Standard 10 and Pro 25 active members, owner included. A server with no plan keeps 3. When a plan goes down or ends, no member is removed, but no new member can join until there is a free seat.
A hosted server stops 15 to 20 minutes after its last use and keeps its data. It starts again when you connect to it, and a few minutes before its next scheduled routine. An open app that sends no request or message for 1 hour does not keep the server on.

Improved

Product analytics records which tools, catalog plugins, websites and routines the agents use, so we can learn which tasks OpenBot is used for. A website is sent only as its registrable domain, such as linkedin.com, never as a URL, path or page title. IP addresses and names with no public suffix, such as .local hosts, are not sent. The name, address and command of a server that you added are not sent. Once a day, the app also sends counts of your agents, routines, skills and servers, and the names of the catalog plugins and curated listings that you use. You can turn off product analytics in Settings → General → Privacy. See PRIVACY.md.
The browser client shows the account, usage and settings controls at the bottom of the sidebar, as the desktop app does. Before, usage and settings were only in the account menu.
The account service sends product analytics events when a plan or a hosted server changes. The events have your account ID and fixed values only, with no email, name or server ID. See PRIVACY.md.

Fixed

A routine runs one time after the computer wakes from sleep. Before, each time missed during sleep started its own run, one after another, and the first run often failed because the network was not back yet. Routines now wait for the network after a wake, and an agent routine does not add a new scheduled run while its previous run is still in the queue.
Close search with Escape or a click outside the search panel. Before, only Command K or a result closed it.
Keep the search tabs in the same place when the result list changes height. Before, one click on a tab could also open the agent that moved under the pointer.
In the desktop app, the live view of a remote host's browser now ends cleanly when the host sends a frame that it cannot show. Before, this caused a main-process error, and the live view did not stop.
The live view of a remote host's browser now ends when its connection closes abnormally. Before, the view could stay open with no new frames, and on the host the end of a live view could also stop a remote desktop session that was running at the same time.
The desktop app now releases the host's live view session when it cannot open the view. Before, the session stayed open on the host and counted against its limit.
The chat scrolls with the mouse wheel or trackpad when the pointer is on a code block. Before, the chat did not scroll until the pointer left the block.
In "Providers & permissions", the "Try it free" arrow points at the OpenCode row. Before, its position came from the screen layout, so it could point at a different row. The arrows of "Try it free" and "Not in the list?" now point at the middle of their targets on both setup screens.
Install updates on Linux again. Before, OpenBot always refused the install with "Another OpenBot session is still running from this application", also when only one session was open.
OpenCode agents no longer stop with "Internal error: OpenCode service failure" after a reconnect or a restart. OpenCode sends this error when it cannot find a stored session. OpenBot now reads OpenCode's session list. If the list does not hold the session, OpenBot starts a new OpenCode session with the conversation history. In all other cases OpenBot keeps the session, tries one more time, and then tells you to try again or reconnect OpenCode.
Start the server list at the top of the web app. Before, the web app kept the empty space that the desktop app keeps for the window controls.
In the agent avatar editor, the "Generated face" heading stays on one line. "Reset to ID" and "New set" are now icon buttons with a tooltip. Before, the heading went onto two lines.
Play the completion sound in Safari in the web app. The sound now also follows the server's mute and notification level.
Remote control connects at once to a computer or server that restarted with no clean disconnect. Before, it waited until the old connection timed out.

Sep 28, 2026

Yaji 0.24.0

Action needed after you upgrade

If you installed the Figma plugin before this version, Plugins shows it as not installed. Remove the old `figma` server from the MCP servers of the agent, then install Figma again from Plugins.

New

Ask an agent to set up another agent after it creates it. The agent can read another agent's profile, model, settings, skills, routines and MCP servers. It can install, turn off and remove local skills for another agent. It can set another agent to Workspace only, turn Computer Use off and change notifications. Only you can give Full access, turn Computer Use on, change auto-approve or change MCP servers. A new agent gets the access limits of the agent that creates it.
Ask an agent to change the provider, model or reasoning effort of another agent, as it can change a name. The agent gets an error that names the available models when a model is not available. OpenBot records which agent made the change, and the previous and new model.
Find local model servers, such as Ollama and LM Studio, and known ACP agents on your computer. Settings shows them under "Found on this computer", where you can add or hide each one. First run shows the same list. The scan sends no key and does not start a program that it finds. You can turn it off, or add more addresses and folders, in Settings.
Add a custom ACP agent, such as Goose, Qwen Code, Cursor or GitHub Copilot CLI, as a provider. "Check agent" starts the command once to make sure that it answers. Its environment variables are stored encrypted on your computer.
Edit a saved custom endpoint. An empty key field keeps the saved key. "Find again" loads the model list from the server.
Leave a joined server from its Server settings. OpenBot asks you to confirm, then removes the server from your server list. The server and its other members stay. To join again, you need a new invitation.
The iPhone app is in public beta on TestFlight. A card at the bottom of the sidebar shows the install steps and a QR code of the invite link. You can close the card. Settings > Mobile Connect always shows the same steps, the QR code and a button that copies the link.
Add Paper to the Plugins tab. Agents can read and change the file that is open in Paper Desktop. Install Paper Desktop, open it once, and open a file. Paper needs no key. The command of a local MCP server can now start with ~/, which is your home folder.
Search is available on mobile. It finds agents and the messages in their chats on the connected computer.
Resize the sidebar in OpenBot web, and drag it narrow to make it compact. A narrow browser window makes the sidebar compact by itself, as in the desktop app.
Open the usage report of a host from its menu in the OpenBot web server rail.
Press Cmd+K or Ctrl+K in OpenBot web to search all conversations on the host.
The "Waiting for replies" block now has a close button when every teammate has replied or failed. Before, a failed request kept the block above the message box until you sent a new message.
Open the server menu from the server name in the web client. The menu switches and adds servers, opens server settings and the marketplace, and sets the rail or menu layout, as in the desktop app.
Show the "Create your first agent" row in the web client sidebar when a server has no agents.
In the web app, change your display name and profile photo and disconnect account sessions. Open Profile from the account menu; it opens in the right panel of the agent on screen.
The web account menu now has usage, Settings (the host's settings), Marketplace, Send feedback, and Message, as in the desktop app. Before, it had only Sign out.
Open an invitation link in OpenBot web. The invitation page has an Open in browser button, and the web client opens the join dialog with the link filled in.
Open a plugin listing in OpenBot web. The plugin page has an Open in browser button, and the web client opens the marketplace on that listing.
"New chat" in the agent settings starts a new chat with the agent. The agent forgets the earlier chat. The earlier messages stay visible above a divider. The instructions, model, tools, memories, workspace, and browser do not change. It is available on desktop and in the web client, also for an agent on a joined server that is on this version or later.
OpenBot now has a macOS build for Intel Macs: download OpenBot-<version>-x64.dmg from the release. It includes the provider runtimes, Computer Use, voice input and remote desktop hosting. Intel Macs get updates for the Intel build, and Apple silicon Macs continue to get the ARM64 build. The Host package (PKG) is still for Apple silicon only.
The web client sidebar has channels, as in the desktop app. You can create, edit and delete a channel, open deleted channels, and use channel memories and routines. The "+" menu and the sidebar context menu have New channel.

Improved

The iPhone app shows an agent plan as a task list, as desktop does: a card that opens and closes, with the state of each step and a spinner on the step that runs. The last message of a chat and its read state do not include plans. Team API v4 now sends the plan of a turn beside its checklist text; older clients and hosts keep reading the text.
The iPhone app shows a file name in inline code, such as package.json, as desktop does: a type badge and the name in the colour of its file type. You cannot open the file from the phone yet.
On the iPhone app, the question form of an agent has no text field of its own. It shows "Reply in the chat" as one more answer row. Tap it, and the next message from the composer is the answer. Without that tap, the composer sends normal messages, also in a channel. A private answer still uses its own masked field.
On the phone, an answer from a teammate agent no longer shows as a queued message that you can edit, steer or move. The queue sheet shows it under "Waiting for replies". A computer on an older version does not send this information, so its answers still show as queued messages.
Show a "Connect your computer" screen in the web app when no computer is connected. It gives the three setup steps and the Download, Join with invitation and Refresh hosts buttons. Before, a small notice showed above an empty conversation. When the computers cannot load, an error message now shows at the bottom of the screen.
Show a centered screen with the same design when your computer is disconnected or connecting. It gives the reason and a Reconnect button. Before, a small notice showed above the conversation.
The web app and a joined computer now show Gemini and custom ACP agents from the host in the model picker, and can start agents on them. An owner or admin can download Gemini on the host. The host and the client must both have this version.
The web model picker shows the host's download progress, the Download and Cancel buttons, and the host's custom endpoints on the Custom tab, for an owner or admin. Your draft stays when you reconnect.
Show the OpenBot logo and an animated loading bar on the sign-in background while the web app loads. Before, the page showed only the text "Loading OpenBot…".
Ask before an import adds an agent whose name is already on the server, for example when you import the same export again. The dialog names the agents and says that each copy is a separate agent. Before, only an info icon next to the name showed this.
Mobile search shows one list of agents and messages, with no filter, in a rounded search field.
The search and add buttons on the mobile home screen are two separate buttons.
The mobile search sheet uses the same background and header fade as the other sheets.
In "Waiting for replies", the avatar of a working teammate no longer moves. The spinner at the end of the row shows that it works.
Agents ask a teammate, or try their connected plugins, the browser and Computer Use, before they say that they cannot do a task. When a site needs a sign-in, the agent asks you to sign in in the browser instead of stopping. When a teammate helps, the answer starts with its name. When nothing works, the agent says what it tried and what you can do to unblock it, such as installing a plugin.
You can make the Dynamic Island as narrow as 20% of its default width. Before, the smallest width was 70%.
The OpenBot logo and the greeting in the idle Dynamic Island keep a 16px inset from the edges. Only the space between them changes with the width.
The web client now opens the first-agent form when a computer has no agents, with the same text as the desktop app and no Cancel button. The model of a new agent is the same default that the desktop app selects.
A host can now have up to 3 active members: the owner and 2 other people. A new join or a reactivation that goes over this limit fails with an error. A host that already has more members keeps all of them. Removed and disabled members do not count.
Server members on desktop and mobile shows how many members the server has of its limit, for example "1 of 3 members". When the server is full, you cannot create an invitation until you remove a member.

Fixed

The web app no longer shows the agent's working notes as chat bubbles. Before, each step of a turn showed as its own bubble, and some of these bubbles were empty. The web app now shows the same messages as the desktop app.
Stop "Provider error" messages for Antigravity info and warning log lines, such as "Checkpoint summary was too long". Before, one message showed at each step of a long conversation. These lines now go to the log. Antigravity error lines still show.
When Codex ignores an unknown setting in its configuration, show one warning that names each setting. Before, a "Provider error" showed only "Codex is ignoring 1 unrecognized configuration setting", with no setting name, and it came back after each reconnect. Codex continues to work.
Create one routine when you ask an agent to make a new agent with a schedule. Before, both agents could save the same routine. An agent can no longer add a routine with the name of an existing one; it changes that routine instead.
Start a new agent on the model that you saved in setup, also when that model is a ChatGPT model. Before, the agent stayed on GPT-6 Luna. Agents from a template, the marketplace or an imported file now also start on the provider and model that you saved in setup. Before, they always started on ChatGPT. When you saved no model, a new agent starts on GPT-6 Luna, and on GPT-6 Luna in the web app too. Existing agents keep their model.
Show the highlight of the selected server across the full row in the mobile server list. Before, the highlight stopped at the end of the server name.
Keep the reason when an agent turn fails. The queue now shows the provider error after the banner closes and after a restart. Before, a failed message kept no reason.
Do not ask a joined server for its MCP server list when you are a member without the admin role. Before, each visit logged a refused request.
Show the loading placeholder in a browser tab preview while the page loads. Before, the preview could show a failure icon until the next capture.
Reload the web app one time when an update removed the files that it needs. Before, the first open after an update could show "This page could not load".
Import an agent when one of its files is already in the workspace, for example two files whose names differ only in case on macOS. The import keeps both files, saves the second one as name (2).ext, and shows a warning. A file that it cannot write is skipped with a warning. Before, the agent was not imported.
Connect the Figma plugin to the MCP server in the Figma desktop app on your computer. Before, the connect step always failed, because Figma does not accept a browser sign-in from OpenBot. The connect dialog shows the steps that turn on the server in Figma. The server can only read designs for now.
On Linux, the window no longer closes a short time after it opens on some GPU drivers. OpenBot now uses software rendering on Linux.
A finished agent reply no longer shows the text of another message after the chat scrolls or loads more messages.
Phones and other devices connect to your computer again when its network comes back. Before, the computer stayed offline for them after Wi‑Fi was off, until you restarted OpenBot.
Show an error and a Retry button on the web sign-in screen when the session check fails. Before, the screen showed the email form with no error.
Stop the web sign-out from staying on "Signing out…" when sign-out fails. The account menu now shows the error, and you can try again.
Show a web session-check error as a notification. Before, the message pushed the dock and the composer out of the window.
Use the plain sign-in background on the web, the same as the desktop app.
Mobile search no longer shows sample English results that did not come from your computer.
Mobile chat previews and search results show message text without Markdown marks such as **.
A tap anywhere on the mobile search field opens the keyboard. Before, a tap near the edge did nothing.
A chat opened from mobile search closes with the same zoom into its row as a chat opened from the list.
Show the full compatibility screen in OpenBot web when a host cannot talk to this version. Before, the web showed "Your computer is disconnected".
Keep the server rail of OpenBot web as wide as in the desktop app. Before, it was 8 px narrower, so the sidebar and the chat moved.
Keep the message that you pick in global search on screen. Before, a conversation that loaded at the same time could scroll to the latest message.
In the web client, files that you drop on a chat or channel, or paste into its message box, now attach to the message. Before, the web client ignored them.
In the browser, a shared or workspace file link in a message opens the file preview. Before, it showed a "desktop only" error.
In the browser, a sent image no longer shows "File not found". Its card shows the name and size, and a click opens the preview.
In the browser, "Open" on a file card opens the preview. Before, "Open" and "Download" both downloaded the file.
In the browser, an image in the composer shows as a file with its name instead of a broken thumbnail.
In the web client, a chat with an agent now shows the unread divider and the new messages banner, and marks messages read on the host. Before, the web client did not show unread messages and did not mark them read.
The web account dock shows your name and email on every browser. Before, it used the macOS layout on every computer.
The web usage indicator reads usage again when a provider connects or disconnects on the host.
A channel now shows a stop button while its agents work. Before, you could not stop a channel run.
On a MacBook with no notch, the Dynamic Island gets narrower when you make its width smaller. Before, it kept an empty space for a notch that was not there, and the width setting had almost no effect.
A narrow Dynamic Island no longer cuts off the greeting at its right edge.
Pause the browser view in OpenBot web while the marketplace or a notification covers it, as the desktop app does.
The web client now shows the queued messages of an agent. You can steer, cancel, reorder, and edit them, as on desktop. The "Waiting for replies" count on the web now agrees with desktop.
Keep the server order that you drag in the web client's server rail. Before, the rail said the server moved, but the order did not change.
Show the real connection state of each server in the web client. Before, every server that was not open showed as offline.
Show server logos in the web client's server rail, server menu and server settings.
The browser preview no longer waits behind the agent's browser actions. While the agent works on a tab, or when a capture is slow, the preview shows the last frame of the same page. Before, a preview on a heavy web app could take up to 30 seconds, and it also delayed the agent's next action.
Browser snapshots are smaller. The page text is limited to 20,000 characters, with the text in the viewport first, and the snapshot says when it left text or elements out. Each snapshot also shows fewer and shorter console and action entries. Before, each browser action could add up to 1 MB of page text and log entries to the agent's context.
On Linux and macOS, OpenBot finds a provider CLI that you installed yourself, such as OpenCode under nvm, when your shell profile prints text at start. Before, a greeting or a tool such as fastfetch in .bashrc or .zshrc made OpenBot show the provider as not downloaded. MCP server commands had the same fault.
The web client now sets the page language to the language of its text. Before, the page always said English, and a screen reader could read other languages with an English voice.
While the web client loads, it shows the OpenBot loading screen from the first frame. Before, the first frame was an empty page.
The web client no longer uses the smooth scroll and the minimum page width of the site pages.
A left-click now closes a sidebar context menu. Before, the menu stayed open when you clicked the empty sidebar area or the row that opened it.
The $ menu in the web client shows the agent's skills, and MCP servers for an owner or admin. Skill tags in messages show as skill chips, and copied messages use the current skill names. Before, the menu was always empty and each skill tag showed as an unavailable skill.
The web client keeps pinned agents and channels, and collapsed sections, after a reload. Before, it forgot them, and only agents could be pinned.
The fullscreen browser in the web client has back, forward, reload, the address field and tab close, as in the desktop app. Before, its toolbar was empty.
The web browser sidebar shows tab previews. Before, the cards had no preview.
A new tab in the fullscreen web browser no longer fails with "The host could not complete this request".
In the web app, a reply streams into the chat as it arrives. Before, the app read the whole conversation again for each part of the reply.
In the web app, the activity line shows what the agent is doing now, as in the desktop app.
In the web app, an agent error shows its own text above the composer, as in the desktop app. Before, a general "The host reported an error" message replaced it.
In the web app, a message that fails to send shows its error one time, above the composer. Before, the same error also showed as a notification.
The web app shows the usage-limit notice above the message box when the selected model has no usage left. Before, only the account menu showed the usage.
In the web app, a reply to an older message shows the quote of that message. Before, the quote was missing until you loaded the older messages.
In the web app, a question card stays open until it has shown your answers. Before, it could close before the answers appeared.
The web app plays the completion sound when an agent with notifications on finishes a task, as the desktop app does.
The web app tells the host when you are writing to an agent, as the desktop app does.

Sep 27, 2026

Yaji 0.23.0

New

Show the plan of an agent as a task list in the chat. The list changes while the agent works and stays in the history. Codex, Claude, OpenCode and Grok agents send plans. On mobile, the plan shows as a text checklist.
Show a "Waiting for replies" block above the message box when an agent asks its teammates. Each row shows a teammate and its status: asked, working, replied or failed.
Let an agent see the progress of other agents, and stop work that it gave to another agent. An agent can stop only a turn that its own messages started. It cannot stop work from you, a routine, a channel or another agent.
Let an agent create a new agent with the provider, model and reasoning effort that you ask for. An agent can also read the models and reasoning efforts of each provider.
Show the images of an agent message in a gallery when the message has two or more images. Click an image in a chat to open it in a viewer. In the viewer, you can go to the next image and download the image.
Read the changelog on openbot.run/changelog. It shows the changes of each release, and the steps to do after an upgrade. The Releases link in the footer now opens this page.
Compare OpenBot with Grok Bot, Muse, Hermes Agent, OpenClaw, Manus, Claude Cowork and Devin on openbot.run/compare. The Grok Bot article moved from News to this section. The old link opens the new page.
Open News, Guides and Plugins on openbot.run from a navigation menu that shows the latest articles. On a narrow screen, a menu button opens the same sections.
Show the reading progress of each section on news, guides and comparison pages on openbot.run. You can share a link to one section.
Show an animated 404 page on openbot.run, with links to the main sections.

Improved

Let an agent read all the replies of its teammates in one turn. Before, an agent that asked two or more teammates started one turn for each reply, and answered you again after each teammate. A message from you does not wait for the replies.
Show the replies of teammates in the "Waiting for replies" block, not in the message queue.
Move the AI providers to their own Settings tab, after General.
Use the onboarding provider list in the Providers & permissions dialog and in the first setup after an invitation. You can add and manage a custom provider there. On a joined server, the dialog shows only the list.
Show the reasoning efforts as a row of segments in the model picker. The header shows the effort after the model name. A provider that is not available is dim and shows a status dot.
Give the action menus and the model picker a new look. Menus now close with a motion, and a submenu moves in from the menu before it.
Check a hosted-site file path before the upload starts, with the same rule as the account server. A path such as api/…, server/… or node_modules/…, or a path longer than 240 characters, now shows a translated error.
Show the Grok sign-in text and the skill folder warning in French and Japanese.
Connect to joined servers again at once when the computer wakes from sleep. Before, the app could wait up to one minute, or longer for a host that was offline.
Show images from a joined server again without a new download each time. The app keeps a copy in memory for 10 minutes.

Fixed

Keep a question from an OpenCode agent open until you answer it. Before, the question failed after 60 seconds with "Request timed out", and a later answer went to nobody. When a provider stops waiting for a question, the question now closes.
Start a new agent on another signed-in provider when the default provider has no model. Before, an agent that created a new agent got the error "Grok has no available model." When no provider has a model, the error now tells you where to change the default provider.
Give an OpenCode agent the full result of a Computer Use action. Before, the agent got only a summary, such as the number of windows, and could not select a window or an element.
Show a teammate message that needs no reply as done when an OpenCode agent does not reply. Before, the message showed as failed, and the chat showed an OpenCode error.
Show the cost of the default Codex model in agent usage. Before, the usage panel showed no cost for this model.
Show the new server logo in server settings before you save it. Before, the settings showed the server initials.
Keep the focus in the server name menu. A click outside the menu now only closes it. Before, the click also went to the page below.
Generate the profile of a Gemini agent without its MCP servers and tools, as for OpenCode and Grok.
Remove the incomplete copy when a copy of a local agent fails. Before, the incomplete agent stayed.
Keep cached remote attachments in their cache folder. Before, some attachment IDs could write a file outside it.

Sep 26, 2026

Yaji 0.22.0

Action needed after you upgrade

Workspace only is now enforced. An agent with Workspace only can write only in its workspace, the shared folder and the temporary folders. On Windows and Linux, a Grok, OpenCode or Gemini agent with Workspace only does not start. To use it, select Full access in the agent settings. Agents with Full access, which is the default, do not change.

New

Use Gemini with a Google AI Pro or Ultra plan. OpenBot downloads Google's Antigravity server to this computer and signs in through your browser. Members of a joined server do not see Gemini agents.
Select More providers on the onboarding screen to add Gemini or a custom provider.
Turn Computer Use off for one agent in its settings. Existing agents keep Computer Use on.
Let owners and admins of a joined server manage the host from their desktop: agent access, auto approve, skills, shared tables, marketplace agents, provider keys and downloads, custom endpoints, and the server name and logo. Members can only read.
Let owners and admins in the browser client use server settings, providers, the marketplace, agent skills, shared tables, auto approve, Try skill, plugin prompts and shared agents.
Update an agent that was added from a marketplace listing to the current version of the listing, from the web client or from a joined desktop.
Show the OpenCode Go quota in Usage when you save a Go key.
Show the servers in a menu instead of the server rail. Select the layout in the server name menu. OpenBot keeps the layout for each device.
Select the interface language on mobile.

Improved

Show more of the interface in French and Japanese.
Install the latest provider CLI on a first download. If the release check is slow or fails, OpenBot installs the pinned version.
Fade streamed text as a trail, and keep the fade across Markdown blocks.
Show the server rail and the top row on one surface.
Hide saved secret values in all logs, and write a local trace line when the main process fails.

Fixed

End the turns of a provider client that stopped. Before, these turns stayed in progress until OpenBot restarted.
Keep queued messages when a file delete fails.
Show a request timeout from any provider as a timeout.
Sign out and revoke sessions correctly while remote sessions are open. Before, this failed with an error and the browser kept its session.
Stop an error after sign-out in the web client.
Move the focus back to the server name after a server row closes the menu.
Correct one Japanese sentence.

Sep 26, 2026

Yaji 0.21.2

Version 0.21.1 was not published because its macOS and Windows builds failed. This version also contains all changes of 0.21.1.

New

Keep a local trace of IPC calls and provider turns in logs/trace.ndjson. Each line has only a time, a name, a duration and an outcome. The file stays on this computer. The diagnostics export adds a summary for each name.
Add a test run to a saved routine on mobile, and show routine events in the mobile chat.
Show the agent name and the latest message when you long-press an agent on mobile.
Change the order of servers on mobile. Long-press a server and select Edit order.

Improved

Retry an offline remote host every 5 minutes while the app has focus, and every 15 minutes without focus. Before, OpenBot retried every minute.
Send sign-in and invitation emails as HTML with a plain-text part. Out-of-office replies no longer answer these emails.
Show a new design for the mobile server list and the agent appearance picker.
Open the mobile photo picker faster, and move and scale an avatar photo in the app.

Fixed

Keep both keys when two providers save a key at the same time.
Keep the stored session when you type a wrong sign-in code for a new challenge.
Close the live browser view when its screencast cannot start again. Before, the view stopped on its last frame.
Stop an error in the sidebar when you switch servers.
Send desktop host analytics events with the app origin. Before, the analytics server refused all of them. When you turn analytics off, OpenBot does not send these events.

Sep 25, 2026

Yaji 0.21.1

New

Publish an agent as a template that other people can add. Press Publish in the agent chat header. OpenBot sends the instructions, the routines and the skills to openbot.run. Files and memories are not sent, and a field that looks like a secret stops the publish. The link opens a page at openbot.run/agents/<id>. Add to OpenBot shows the template in the app, and OpenBot adds the agent only when you click Add agent. Unpublish removes the template. A new publish keeps the same link.
Edit a routine schedule with chips in the routine panel and in the routine card in the chat.

Improved

Show the account profile name and avatar in the browser client.
Show that a sign-in server does not accept OpenBot when it refuses the registration, for example Figma. Before, OpenBot showed a credentials error. The Figma listing tells you to use the local MCP server of the Figma desktop app.
Get faster answers from the account server.

Fixed

Delete an agent on Windows. Before, deletion failed with EBUSY because a provider session kept the agent workspace open.
Remove a deleted agent from its channels. Before, the channel showed an Unavailable member. OpenBot also removes these members at startup.
Start OpenBot on Linux from a terminal. Before, the window could freeze at startup.
Send a steer message with the same framing as a new turn.
Redact each line of remote diagnostic logs.
Show an error when the copy of a channel message or a plugin link fails.
Use the fallback name for a download with a bad file name.

Sep 24, 2026

Yaji 0.21.0

New

Import agents from Grok Bot in Server settings, Import. Add the OpenBot export agent to Grok Bot, send it one message, and choose the .zip file it saves. OpenBot imports each agent you select with its instructions, avatar, skills, routines and memories, and optional workspace files. The file is read on this computer and is not uploaded. Chat history is not copied.
Dictate a message on mobile. With an empty message field, press the microphone. The text shows in the field while you speak, and you send it when you are ready. Set the language in Settings, General, Dictation.
Delete a channel from the channel row menu on mobile. Only server owners and admins see Delete.
Show Skills and Files in the agent info on mobile.

Improved

Use less memory and CPU. OpenBot stops a provider CLI that no agent uses after 60 seconds, keeps fewer idle Claude, OpenCode and Grok sessions open, and streams attachments from disk. The first message to an agent after a long idle time can take a little longer.
Scroll and type with fewer frame drops in the mobile chat and agent list.
Use a darker dark-mode background for the mobile screen, drawer and sheets.
Show the provider Update button on the same row as the update message.
Save an agent with empty instructions on mobile.

Fixed

Remove a question or approval from the screen when its provider stops. Before, it stayed until a restart, and each answer failed.
Keep the approvals of other providers when one provider stops.
Keep pending questions and approvals in the web app when the host sends a partial update.
Type Chinese, Japanese and Korean with an input method in the message field. Before, a Latin letter went in before each segment.
Show the size of a stored file that is a link, not the size of its target, in Storage.

Sep 23, 2026

Yaji 0.20.1

Fixed

Keep an animated WebP avatar animated after upload. Before, OpenBot kept only the first frame.
Hide the secret code input while OpenBot sends the code. The input shows again if the code fails.
Wrap long text in the avatar upload control.

Sep 23, 2026

Yaji 0.20.0

Action needed after you upgrade

Desktop notifications now appear. If you turned off desktop notifications, turn them off again in Settings.

New

Show the skills in an agent's workspace skill folders that OpenBot did not install, such as .agents/skills. OpenBot does not change these folders. A skill shows a problem when its SKILL.md is not valid, or when the agent's provider does not read its folder.
Show disk use and files in a Storage tab in Server Settings, a Files view in agent settings, and a Files panel in the chat. An owner or an admin can delete a file and clear caches and logs.
Mute a server for a set time or until you turn it back on, and set its notification level: all activity, only when it needs you, or nothing.
Send a test notification and open the system notification settings from Settings.
Mark an agent as Workspace only in the agent settings. The setting is not enforced yet.
Check for provider CLI updates from a provider's actions menu. OpenBot offers the latest upstream release of Codex, Claude, OpenCode, and Grok without a new OpenBot version.
Add Linux arm64 builds.
Add a Composio listing that takes your own MCP link.
Show file type, size, previews, and upload progress for attachments on mobile, and show images that an agent generates.

Improved

Accept a skill description of up to 1024 characters, as the Agent Skills specification allows.
Show the open agent's provider in the usage chip.
Apply edited agent instructions to Codex and Claude sessions. The agent keeps its thread.
Ask for confirmation in one shared dialog everywhere.
Hide the query string of an MCP URL in logs, and mask MCP header and environment values in the server MCP panel.
Send fewer requests to the account service. Without app focus, an offline host retries after 15 minutes.

Fixed

Make Apple and Google sign-in work in the embedded browser.
Open OpenBot on a second launch after a quit that did not finish on Windows.
Show new chat replies when reduced motion is on.
Center the idle emoji in the Dynamic Island.

Sep 23, 2026

Yaji 0.19.0

New

Use OpenBot in a web browser at openbot.run/app with your account and your connected hosts.
Set the Dynamic Island width and height in the new Dynamic Island settings tab on macOS.

Improved

Show every model a provider reports, newest first. When a provider no longer lists an agent's model, the agent moves to the provider's default model and keeps its thread.
Use less memory: close a provider CLI after 10 idle minutes and resume its session on the next message, restore only the active browser tab at startup, and keep fewer messages of closed agents in memory.

Fixed

Show one notice when a Codex plan reaches its usage limit, not an extra "could not continue" error.

Sep 23, 2026

Yaji 0.18.0

New

Use OpenBot in French.
Open invitation (including permanent links), pairing, and plugin links directly in the mobile app.
See a running, waiting, blocked, or failed routine as a mark on the agent row.
Update a provider, or sign in with a code, from the new "More actions" menu on each provider row.
Start with OpenCode free models without a sign-in.

Improved

New ChatGPT agents start on GPT-6 Luna, and new Claude agents on Claude Opus 5.5. Existing agents keep their model.
Refresh the server list on other signed-in devices when a membership changes.
Show the routine run history in chat as one summary per run.
Improve mobile reply animations and keyboard scrolling.

Fixed

Send live view clicks to the correct point of the remote page.
Keep Claude narration between tool calls out of the chat bubble.
Keep embedded browser sign-in popups and OAuth sessions working.
Load a remote server's agents when one agent uses a 1M-context Claude model.
Show why an updated OpenCode CLI stopped, instead of "ACP connection closed".
Stop showing a Grok tool error as a provider error.
Stop showing a refused routine as a provider error; the agent now gets the reason and can correct it.
Keep failed Codex background refreshes out of provider error messages.

Sep 21, 2026

Yaji 0.17.0

Action needed after you upgrade

On macOS, grant Screen Recording and Accessibility to OpenBot for Computer Use. Grants for the old Codex helper do not carry over. Remote desktop uses separate Sunshine grants in the macOS account that runs it.

New

Use Computer Use with Codex, Claude, Grok, and OpenCode. OpenBot now includes the Computer Use driver on macOS, Windows, and Linux.
Set up remote desktop access from Server Settings. Check Sunshine permissions, display availability, and the macOS user session; open a helper to grant access.
Test remote desktop video, mouse, and keyboard with a temporary test panel, including a local test on the same Mac. Older runtimes support a video-only test.
Enter passwords and authentication codes through secure browser prompts in chat. Secret values stay out of agent messages and browser captures during submission.

Improved

Enable ordinary auto-approval by default for new agents. Existing saved agent choices are preserved. Automatic website publishing, replacement, and deletion require Turbo.
Keep previous approval settings in their original file during upgrade.
Rename a saved MCP server named computer_use to an available computer_use_saved name. Its configuration and credentials are preserved.
Explain when updates are managed by the host in Settings.

Fixed

Backport Sunshine security fixes for malformed input packets, pairing approval, and exact client certificate checks.
Keep provider downloads available while provider checks run.
Support Canva and other MCP sign-ins that require a local redirect address.
Allow plugin removal from its marketplace page.
Keep local remote desktop tests independent of account ICE settings, reject malformed viewer URLs, and avoid repeated Sunshine restarts while granting permissions.

Removed

Remove the separate Codex Computer Use plugin setup and the Agent CLI setup card.

Sep 21, 2026

Yaji 0.16.0

New

Sign in to ChatGPT with a code typed on another device. OpenBot shows the code, opens the verification page, and reports when the sign-in finishes.
Set standing approval for one agent or enable Turbo mode for all local agents. Turbo keeps each agent's individual approval when it is turned off.

Sep 21, 2026

Yaji 0.15.3

Action needed after you upgrade

A server declared outside OpenBot no longer reaches a Claude or Codex agent. Add it in Settings, MCP to keep it. This covers ~/.claude/settings.json, a project .mcp.json, a plugin, agent frontmatter and ~/.codex/config.toml. Nothing is deleted: the declaration stays in its own file and OpenBot does not read it. A notice says this once. OpenCode and Grok document no equivalent setting, so their agents can still start servers from their own files, which the panel now states.
Sign in again once after upgrading to reach Canva, Figma, Linear, Notion, Sentry or Stripe. The six listings now connect over HTTP and OpenBot holds the OAuth client itself: it opens the browser, keeps the tokens in the operating system's secret storage, and adds the header when it hands the server to a provider. No third-party bridge program is downloaded or started. A bridge that already holds a token keeps it in its own folder, which OpenBot does not read, so the sign-in does not carry over. Removing the server forgets its sign-in, and the tokens are redacted on every log and export path like every other secret.

New

Start an MCP server on a computer that has no Node. OpenBot downloads Bun 1.4.2, a JavaScript runtime it owns and checks, and starts a STDIO server with it. A computer with its own Node keeps using that one: the managed runtime is the floor, not a replacement.
Say why an MCP server did not reach an agent. A command this computer does not have, and a working directory the provider cannot carry, each raise one notice naming the server and the reason, and the panel states the working-directory limit before the server is saved.
Answer a question an MCP server asks, on every provider. A server that needs a field gets one question for each field it requests, with the typed value it asked for, a secret hidden as it is entered, and a decline that names the field the server cannot do without.
Invite people with a permanent link, in the new Perma link tab of server settings. The link is reusable and never expires, so it suits a channel or a document rather than one person. A host keeps at most five, each one is revoked from the same tab, and a permanent link cannot be bound to an email address. Single-use invitations are unchanged. A server that speaks an older Team API does not offer the tab, because that protocol carries no permanent link.
Update OpenBot for every account on a shared Mac from one place, with the new Host package. An administrator installs it, and the Mac then downloads a release once, waits until every signed-in account is idle, installs it one time, and starts OpenBot again for each account. A managed account shows the host's progress and never installs on its own. The package is signed and notarized on its own, and it installs managed-host infrastructure only: the DMG is still the application. The installer also creates Standard accounts for tenants.

Improved

Replace a Codex thread that started before this release once, keeping its conversation and its history, so the tools the agent has match the panel.
Re-read a provider's remaining usage while the figure is on screen, so a window that counts down no longer reads minutes out of date. The dock, the account menu and the usage popover refresh a reading that is five minutes old. Nothing is read while no view shows the figure, and a hidden window reads again when it comes back.
Run two Remote Desktop sessions at the same time on a shared Mac. Each signed-in account gets its own port range, so one session no longer takes the ports the other needs.

Sep 19, 2026

Yaji 0.14.1

Improved

Sign in to Claude from the composer notice, which opens the OAuth window instead of the authentication documentation. OpenCode, whose key is pasted in settings, no longer raises a notice with a button that starts nothing.
Name the provider under each mark in the header model picker rail, with its state below the name.

Fixed

Ship the 0.14.0 work, which no build carried: that release stopped while it was being packaged, on a test that allowed one second for the interface to answer on a machine that was building and signing the application at the same time. Everything listed under 0.14.0, the fourteen-plugin Marketplace catalog included, arrives here.

Sep 19, 2026

Yaji 0.14.0

New

Show a host's browser tab live to a remote member, with pointer and key input back to the page.
Connect an MCP server before it is installed: the install dialog tests the configuration, and only a configuration that answers is saved.
Open a Marketplace plugin from an openbot://plugins/<slug> link on its own detail page.
Publish plugin pages on openbot.run from the shipped catalog: fourteen listings, each with its own mark.
Give each agent one silhouette and a face that matches the work, in the sidebar and the activity indicator.
List Marketplace plugins in the landing header on openbot.run.

Improved

Define when an agent's browser tabs close: no tab closes when a turn ends, the agent uses close_tab when a task no longer needs the tab, and deleting an agent closes its tabs.
Show the download percent while a provider connects, with Cancel during the download and Retry on failure.
Grow the plugin catalog to fourteen simple-auth listings: GitHub, Linear, Notion, Figma, Sentry, Context7, Stripe, PostHog, Airtable, Firecrawl, Brave Search, Resend.
Name the shared tables "Tables" in agent settings.

Fixed

Let the embedded browser copy: pages can write to the clipboard on user gesture, and right-click offers copy link, copy image address, cut, copy, paste, and select all with full link targets.
Align the Usage header icon with the provider logos.

Sep 18, 2026

Yaji 0.13.0

New

Install an app and its skills together from the new Marketplace Plugins tab. Each plugin adds its MCP server record and the pinned skill versions its instructions need.
Let agents keep records in one shared database at ~/OpenBot/Shared/Data/agent-data.db. An agent creates, reads and writes its own tables, and the Shared tables view lists every table with its owner and lets you delete any of them.
Let an agent tell a teammate something without asking for an answer. The message names that no reply is due, so the teammate does not open a turn for it.
Let the agent complete provider authorization steps itself instead of stopping for the user.
Show the remaining usage for every connected provider, and show the Grok account email.

Improved

Handle an exhausted provider usage limit: the agent reports the limit and the time it resets instead of failing the turn.
Open a browser takeover page from its preview card. A takeover request no longer expands the browser over the conversation on its own.
Open an attached file in the right panel instead of a modal.
Group the agent chat message times, and correct the position of a message timestamp.
Use a custom agent avatar in the activity indicator.
Remove the View source toggle from a Markdown preview.

Fixed

Drop the placeholder answer that an agent sent for a teammate request.
Load an OpenCode ACP session before reading it at startup.
Remove the duplicated agent message previews.
Save an edited agent instruction again.
Keep a required input prompt visible.
Type into the focused page when the browser cannot target an element.
Keep the agent recipient menu text readable.

Sep 17, 2026

Yaji 0.12.0

New

Open a Marketplace Skill or Agent on its own detail page, with a crumb back to the listing, a debounced search, and one install control that picks the target agent.
Show category artwork on every catalog Skill instead of the generic fallback glyph.
Preview Markdown and XLSX spreadsheet attachments in the file preview panel.
Create a channel from the sidebar context menus and the sidebar topbar.
Name the missing screen recording grant in Remote Control, open System Settings from the host's own Server settings, and re-read the grant with Check again.
Show a themed splash backdrop on mobile startup.

Improved

Show the expanded browser edge to edge. Leave it with the button in the top right corner or with Escape, while a text field in the page keeps Escape.
Reduce the hover area of the compact Dynamic Island.
Respect per-server mute on the notch.

Fixed

Hide the mobile loader that stayed over the chat.
Use the shared provider names on mobile and truncate long model labels.
Hold the mobile splash until the artwork shows on a fast startup.

Sep 16, 2026

Yaji 0.11.0

New

Edit a queued message before the agent starts it, on desktop, mobile, and over Team API.
Download one file straight from a message with the per-file Download action.

Fixed

Keep the original file name on downloads instead of a generated one.
Sign in to Google inside the embedded browser.
Open WhatsApp Web login instead of the unsupported-browser page.

Improved

Hide sidebar row time and date when the panel is narrow.

Sep 15, 2026

Yaji 0.10.1

New

Download all files from a message with three or more attachments as one ZIP file.
Enter an OpenCode key for paid OpenCode models, with refresh and status in the provider row.
Choose provider, model, and reasoning effort when you create an agent, where the host supports it.
Read the WTF Is OpenBot guide under /guides.
Open Settings from the application menu with a keyboard shortcut.

Improved

Show provider errors on one composer card instead of in the transcript.
Write only the streamed message on each flush, and cut idle CPU and per-frame render work.
Anchor the stopped task banner above the composer.
Fix dismissible chat-scoped error banners so they stay dismissed in their chat.

Fixed

Fix OpenCode model list, agent setup picker, permissions text, and chat copy.
Fix folder listening setup that failed on routine interval validation.
Handle dynamic channel tools with no active assignment.

Sep 15, 2026

Yaji 0.10.0

New

Manage MCP servers from server settings, including save, remove, enable, and test actions.
Add mobile channels, agent pins, and workspace connectivity, with channel chat, records, and task actions. Preserve deleted channels as read-only previews.
Add custom agent photos to the mobile avatar experience.
Queue a message to a busy agent and show why it waits.
Add an OpenCode download action to the custom provider row.
Add an Introducing OpenBot article under /news.

Improved

Repository moved to nightly-labs/openbot; release links and the update feed follow it.
Tighten agent-to-agent communication prompts and route agent instructions instead of front-loading them.
Stop the two animations that burn idle CPU, and add dev:cpu to measure it.
Explain a failed update check, make the retry answer, and stop routing updates through a rename redirect.
Display channel titles in channel UI; draw a routing receipt as channel activity, not a message.
Fix mobile sheet sizing, restore progressive blur on iOS, and update the iOS app icon asset.
Fix browser preview layout, tab selection, address navigation, and sidebar icon styles.
Keep chat input scroll position while editing earlier lines.

Sep 14, 2026

Yaji 0.9.0

New

Run OpenBot on Linux x64. The release publishes an AppImage, the one Linux format that updates itself in place, and the in-app provider download covers Codex, Claude, Grok, and OpenCode there. Voice prompts and remote desktop are absent on Linux and report themselves as unavailable, and the AppImage is unsigned, as the Windows installer is. docs/TROUBLESHOOTING.md covers the AppArmor profile that Ubuntu 23.10 and later need.
Give an agent skills. Write a skill locally, install one from the marketplace, and enable, disable, or remove it for each agent. Type $ in the composer to insert a skill, and read a shared chat's skill actions in its preview. @ still means agents and files.
Choose the language of the application. Settings offers System, English, and Japanese, the menu and agent notifications follow the choice at once, and no restart is needed. A key that is not translated yet reads English rather than disappearing. The Japanese catalog is a first draft.
Preview audio, video, SVG, and email files in the file panel and the attachment lightbox. Audio and video keep previewKind: "none" on the wire, so no released Team API adapter changes meaning.
Read /news and /guides on openbot.run: article pages, an RSS feed, a sitemap, and structured data, with article artwork drawn from the title.
Say what OpenBot costs on the landing page.

Improved

Start a new agent on low reasoning effort. Codex CLI reports medium for every GPT-5.6 model, which buys little on GPT-5.6 Luna, the model a new agent starts on, and costs a wait on every turn. An agent whose effort you already set is not moved, and every effort stays available in the picker.
Rework the composer mention and skill picker. It hangs off the input, wears the queue panel's colors, names the type of each option and the source of each skill, and travels to its new height as the query narrows. The queue panel gives up that space while the picker is open.
Drag the agent sidebar down to 128px before it snaps to the avatar rail. A width you already set stays inside the new range.
Keep one provider CLI store for each computer, under provider-runtimes beside the application data. The packaged application already used that path, so nothing moves for an installed user. Several OpenBot instances can now install a provider at the same time without taking a running CLI away from each other. Development profiles no longer offer the same update after every restart.
Move switches and camera panels with their own motion, and animate an article background continuously.
openbot.run now reports the five UTM tags on a campaign link (utm_source, utm_medium, utm_campaign, utm_content, utm_term), which the previous privacy note said were never transmitted. Every other query parameter and the hash are still dropped, so an invitation token cannot reach analytics. This is the public website only: the desktop application is unchanged, and chats, files, and commands are never sent. PRIVACY.md and ANALYTICS.md describe what an event carries.

Fixed

Stop Grok's telemetry export failure from showing as a provider error. A computer that cannot reach an OpenTelemetry collector raised a "Provider error" toast on every switch to Grok. The CLI colour codes it printed as text are gone too, and a colour sequence can no longer hide a secret from the redactor.
State a signed-out or spent provider above the composer. A lapsed account answered with its whole HTTP exchange, and the account menu printed the URL, the headers, and the 401 body. The composer now offers Sign in before you send, or names the reset time on a spent plan window.
Keep undo working for composer typing, and stop a character being written twice.
Play an attached recording. The packaged application blocked audio and video playback, and an attached .eml file could not be previewed at all.
Retry the QR scanner after the camera fails to start.

Sep 11, 2026

Yaji 0.8.0

Action needed after you upgrade

On mobile, OpenBot now sends product analytics, and the setting starts on. Turn it off in Settings at any time; events wait until you sign in and are dropped if you refuse. Nothing on the desktop app changed, and chats, files, and commands are never sent. PRIVACY.md describes what an event carries.

New

Add channels: a shared chat that several agents join, coordinate a task in, and answer in together. A channel carries its own title, instructions, memories, and routines, and it pins to the sidebar the way an agent does. Channels travel over the Team API as channel-chats-v1, so a phone or a joined server on an older build keeps working without them.
Describe your own OpenAI-compatible endpoint and pick its models anywhere a model is chosen. OpenBot encrypts the API key on this computer and gives it only to the local OpenCode process, and keeps it out of every export, log, and diagnostics report. Every saved endpoint sits under one Custom provider row, because which endpoint an agent uses is a model choice rather than a provider choice.
Preview open browser tabs in a sidebar, and expand one into a full-width panel.
Mute desktop notifications for one server without muting the others.
Switch servers with numbered keyboard shortcuts.
Count the new messages on the chat scroll button, so a thread that moved while you read tells you how far behind you are.
Download and pin the OpenCode CLI, like the other three providers. The Install button that sent you to the OpenCode website is gone.
Use OpenCode's free models with no account and no sign-in. A new OpenCode agent answers as soon as the CLI is downloaded.
Add an optional OpenCode Zen key in Settings for the paid OpenCode Zen models. OpenBot encrypts the key on this computer, gives it only to the local OpenCode CLI, and keeps it out of every export, log, and diagnostics report.
On mobile: open an agent's information and edit it in a sheet, reply to a message and use the message actions, send attachments from files or the camera, and pin an agent with a full swipe that a screen reader can also do.
Choose whether mobile gives haptic feedback.

Improved

Keep an OpenCode CLI you installed yourself. OpenBot reports its version, offers an update, and never forces the download.
Leave the OpenCode Go models out of the model picker when a Zen key is what lists them. OpenCode reports Zen and Go as one catalog, but a Zen key does not buy Go, so those models answered every prompt with "Invalid API key.". They still appear when you signed in to Go in OpenCode itself.
Start a new OpenCode agent on a free model, Muse for choice. OpenCode reports the services you signed in to before its own, so a new agent picked a model behind one of those sign-ins and its first message could fail with "Token refresh failed: 401" while the free models sat further down the list. No model that bills is ever the default now.
Create an agent without instructions. The field was required for no reason a user could act on.
Give every agent interaction card one shape across the app.
Load mobile chat history as you scroll, instead of holding a whole thread in memory.
Move the mobile save actions into the native sheet headers, and show pinned agents in a grid with a stated capacity.
Raise the text contrast of the mobile theme.

Fixed

Show OpenCode tool activity again, and stop hiding browser actions.
Keep channel reads, signed-out authorship, and sidebar order correct.
Preserve Claude history answers, and render a resolved question as resolved.
Let a remote agent avatar download over WebRTC. A paired phone asked a Team API v2 host for an avatar with no request body, and the host refused the route.
Accept a custom avatar file an agent names in a prompt.
Mute embedded browser tabs by default, and stop counting another agent's tabs against your tab limit.
Show local dates on older chat messages and older sidebar chats.
Remove the size label from the embedded browser.
Make the desktop agent purpose optional, as the form already implied.
Recover a mobile session and a server connection after the app loses one, and put the mobile chat keyboard back where it belongs after it is dismissed.

Sep 9, 2026

Yaji 0.7.0

Action needed after you upgrade

A paired phone can no longer disconnect a desktop session. Sign out or revoke a desktop session from Settings on the computer that runs it. Phones and other sessions are unaffected, and no session is ended by this update.

New

Add OpenCode as a fourth provider. OpenBot drives your own installed OpenCode CLI over ACP, so it is not downloaded or pinned like the other runtimes: install OpenCode and run opencode auth login, then pick it when you create or edit an agent.
Add Team API v4, which carries the new provider and agent duplication. Version 1 to 3 stay registered and unchanged, so a phone or a joined server on an older build keeps working on the protocol it already speaks.
Search models, grouped by provider and reasoning variant, in the model picker.
Mention an agent in a mobile chat, see the other participants of an exchange, copy a highlighted code block, and retry a send that failed.
Scan a pairing QR code from an inline sheet on the mobile sign-in screen.

Improved

Give every provider one name across the app. Initial setup and the thread status said "Codex" where the picker said "ChatGPT". All of them now say "ChatGPT".
Ask the user to wait, with a countdown, when the mail provider refuses a sign-in code or a team invitation because the sending mailbox is over its quota. A refusal the sender cannot wait out, such as a full recipient mailbox, stays a delivery failure.
Install and activate OpenBot's pinned provider CLI update instead of asking the CLI to update itself. A CLI update waits while a provider sign-in is pending.
Start the mobile app on the connected route, and block interaction while it loads.

Fixed

Report a provider CLI update that OpenBot refuses to start, instead of leaving the offer on screen.
Keep an agent whose stored profile holds a value this release cannot read, instead of refusing to start. The startup error now names the field it cannot read.
Keep an empty sidebar section visible when it holds no agents.
Keep a chat-created agent in the section of the agent that asked for it.
Prefer an agent in the same sidebar section when one agent messages another.
Remember the selected agent for each server.
Focus Delete in sidebar confirmation dialogs.
Report an OpenCode turn that produced nothing as a failure, instead of a blank reply.
Desynchronize desktop avatar animations, animate idle sidebar agents, and keep an avatar pose across a change of agent activity.
Keep the Dynamic Island unfocusable while the mouse is over it.
Explain what to do when a file preview or another surface fails, on desktop, mobile and the web.
Keep a mobile profile name when the name field is left blank.
Remove the scroll edge effect from the mobile add-server sheet.

Sep 8, 2026

Yaji 0.6.1

Fixed

Fix Usage reports failing to load in the desktop app because Electron could not clone the request.

Sep 8, 2026

Yaji 0.6.0

New

Add agent and host-wide usage reports with model, provider and daily cost details.
Add mobile server settings, member management and invite QR flows.
Add marketplace presentation controls for agent categories and creator avatars.
Show all integrated provider models and let users update a provider CLI from OpenBot.

Improved

Check for desktop updates every four minutes while OpenBot is running.
Preserve mobile sign-out when offline by deferring account revocation.

Sep 7, 2026

Yaji 0.5.0

Action needed after you upgrade

Paired phones must be unpaired and paired again after this update. A phone stores the agent identifiers it was given, and those identifiers have changed, so its pins and unread markers point at agents the host no longer knows. Chats open normally again once the phone is re-paired; nothing on the computer is lost.

New

Connect a phone to OpenBot. Pair a device from the desktop app, then read and answer agent chats from the mobile app, with connection status on every server surface and profile changes synced to each connected device.
Review and revoke account sessions and paired devices from Settings. A revoked credential ends the remote sessions that used it immediately; other devices stay connected. Signing in again restores the device.
Browser Automation V2: agents drive the built-in browser through native Chrome DevTools Protocol control instead of injected scripts.
Create an agent through a guided setup with a customizable avatar, and manage agents and sidebar sections by asking in a conversation.
Tag an agent or a skill directly in a chat message.
Show detailed live agent activity, including the provider's reasoning.
Import EML files as chat attachments.

Improved

Retheme the desktop app onto one colour, type and icon system.
Rename an agent's Description to Instructions, and grow the field with its contents.
Remove the limit on sidebar agent pins.
Isolate the local team server per OpenBot account. The single-host file the previous build wrote is imported on first launch and left in place; nothing is deleted.
Update the bundled provider runtimes: Codex to 0.153.4, Claude Code to 2.1.263 (@anthropic-ai/claude-agent-sdk 0.3.263) and Grok to 1.0.22. The runtimes are downloaded on demand, so each one is fetched again the first time you use it after this update; the previous copy stays on disk, and a system-installed CLI of your own is not touched.
Finish naming the product concept agent everywhere: identifiers, IPC channels, CSS, copy, the mobile app, and the instructions and tool parameters the models read. Existing agent identifiers are rewritten from bot-<uuid> to agent-<uuid> and every workspace moves from ~/OpenBot/Bots to ~/OpenBot/Agents on first launch. The Team API wire protocol is unchanged — versions 1 to 3 still spell the agent botId, and a translation layer converts.

Fixed

Discover the available ChatGPT, Claude and Grok models automatically again.
Report weekly usage for the model that is actually active.
Recover a provider session the CLI no longer knows, instead of losing the conversation.
Keep chat table layout, message spacing and composer alignment consistent.
Restore the main window when you activate OpenBot on macOS.
Stop the Dynamic Island from clipping as it collapses, and idle avatars nobody is watching.
Fix sidebar search focus highlighting and drag overlap, and the mobile sidebar swipe over agent rows.

Sep 2, 2026

Yaji 0.4.3

New

Persist and restore the main application window position between launches.

Fixed

Apply the Automatically download updates setting: it is now persisted in the user data directory and drives the updater, instead of being a renderer-only value that reset on every launch and never started a download.
Stop macOS updates hanging on Preparing update…: the restart action is offered as soon as the download completes rather than waiting for a native staging event that never arrived, and the phase that could hang is gone.
Bound every update stage: a check, download, or restart that stops responding now reports an actionable error instead of waiting indefinitely. A failed download is retried in place; a failed install asks for a relaunch, because shutdown preparation cannot be repeated safely.

Sep 1, 2026

Yaji 0.4.2

Improved

Stabilize the signed macOS release gate by running the full repository test suite with bounded Vitest concurrency.

Sep 1, 2026

Yaji 0.4.1

New

Add a self-hosted Bun Signal service, coturn, DNS-01 certificate renewal, and Docker Compose deployment under remote/.
Add WebRTC Team API protocol v2 with RPC, ordered events, binary file transfer, backpressure, integrity checks, and resume support.

Improved

Identify signed-in desktop accounts in OpenPanel with normalized email profile traits and ordered event delivery; keep development builds and localhost analytics-free.
Keep accounts, host configuration, memberships, invitations, logical sessions, and public assets in Cloudflare while remote data uses WebRTC only.
Remove cloudflared from the active host transport and from macOS and Windows application packages.
Require old Team API clients and hosts to update before they can use the retired tunnel endpoints.

Fixed

Keep the embedded browser panel closed until requested and prevent repeated toggles from opening duplicate tabs.
Restore embedded X sign-in with persistent cookie consent, a compatible browser identity, and current login routing.
Flush the embedded browser profile before restarts, system shutdowns, and application updates so authenticated sessions persist.

Aug 30, 2026

Yaji 0.4.0

New

Add a macOS Dynamic Island overlay for messages, questions, approvals, browser takeovers, and failures.
Add persistent question prompt bubbles with safe handling for secret answers.
Let agents attach local response images with preview, download, and file actions.
Negotiate compatible Team API protocol versions and capabilities between clients and hosts.

Improved

Redesign the account dock, usage popover, account menu, and Marketplace access.
Improve browser takeover cards with persistent page previews and completed states.
Hide People communication by default while retaining an explicit opt-in.
Require macOS 13 or newer after the Electron and Chromium upgrade.

Fixed

Support embedded Google sign-in with the current Chromium identity while keeping the OpenBot browser session.
Clear Bot unread state reliably when a chat opens or refreshes.
Preserve agent names before title badges truncate in the sidebar.

Aug 27, 2026

Yaji 0.3.5

Improved

Download and verify the Whisper model on first voice use instead of shipping it in every application update.
Remove provider runtimes from macOS and Windows application packages while keeping system CLIs as the first choice.
Install application updates only after the user selects Restart and install.

New

Download Codex, Claude, and Grok provider runtimes on demand from pinned vendor artifacts.
Show provider download, setup, retry, and connection actions in onboarding, Settings, Bot setup, and model selection.

Fixed

Wait for native macOS update staging before offering a restart, and require an explicit update restart on Windows.
Reject oversized or inconsistent release artifacts and remove duplicate native runtime files from packages.
Stream large provider downloads to disk, support safe resume, and reject unsafe or invalid runtime archives.
Respect Windows shutdown and sign-out without starting the NSIS updater.

Aug 26, 2026

Yaji 0.3.4

Fixed

Verify source-built Windows remote desktop binaries as intentionally unsigned while preserving strict vendor signature checks.

Aug 26, 2026

Yaji 0.3.3

Fixed

Isolate Windows package signature checks from the PowerShell 7 module path used by GitHub Actions.

Aug 26, 2026

Yaji 0.3.2

Fixed

Stage Claude and Grok runtimes on the destination volume before the atomic Windows install switch.

Aug 26, 2026

Yaji 0.3.1

Fixed

Resolve bundled provider and fallback CLI paths with the target platform's path format.

Aug 26, 2026

Yaji 0.3.0

New

Add the Grok CLI provider and bundle the supported provider CLIs for a more reliable onboarding experience.
Add attachments directly in the conversation composer and connect the agent marketplace.
Add per-participant emoji reactions, including optional context-aware bot reactions to user messages.
Add improved server invitation and member workflows.

Improved

Rebuild chat messages on reusable bubble primitives and refresh settings, routines, and related popovers.

Fixed

Preserve the active server across refreshes and make server address copying reliable.
Improve composer mentions, worktree setup, invitation handling, and emoji reaction presentation.

Aug 25, 2026

Yaji 0.2.1

New

Add scheduled agent routines with timezone-aware schedules, manual test runs, run history, and delivery in the agent chat.
Add agent memories, shared sidebar sections, and the skills marketplace.

Improved

Refresh the sidebar, agent settings, dialogs, buttons, switches, selects, and Storybook examples with the shared visual system.

Fixed

Prevent resumed routines from running missed schedules and protect unsaved routine edits before navigation.
Render newly appended chat messages without virtualizer refresh loops.

Aug 24, 2026

Yaji 0.2.0

New

Add the production Create Bot flow with practical suggestions, synchronized animated avatars, and first/additional Bot modes.

Improved

Create a Bot only after its complete profile is submitted, then queue its initial role message as one rollback-safe operation.
Replace the legacy new-agent picker and empty-chat onboarding with the dedicated Bot setup screen.

Fixed

Keep the technical development client out of a normal private dev server while preserving the two-client test harness.
Limit the first visible Bot message to its ongoing role.

Aug 24, 2026

Yaji 0.1.22

New

Add a resizable browser Picture-in-Picture panel that stays visible while the conversation remains usable.

Improved

Keep application and conversation state stable during renderer hot updates, including selections, drafts, search, panel state, and active resources.
Simplify the Remote Control toolbar and show agent animation while a remote desktop connection starts.

Aug 23, 2026

Yaji 0.1.21

Fixed

Restore the Solid signals runtime dependency required by production builds.

Aug 22, 2026

Yaji 0.1.20

Improved

Split team IPC registration, renderer message projection, voice status helpers, and workspace path handling into focused modules.
Remove unused dependencies, renderer exports, preview helpers, and an inactive visual test suite while keeping active Storybook coverage.

Fixed

Hide unexpected Team API failures from remote clients while preserving controlled validation errors.
Bound unauthenticated sign-in rate-limit state and reject oversized WebSocket event frames before application parsing.
Make team and remote-server state writes atomic, isolated, and able to recover after a failed write.

Aug 22, 2026

Yaji 0.1.19

Improved

Load the interactive landing preview from server markup and retry its ready handshake until playback starts.
Cache the verified Whisper model in application CI and release workflows.

Fixed

Exit a second desktop app process immediately when another OpenBot instance already holds the profile lock.
Keep the agent activity avatar visible for 500 ms after streaming ends and preserve its layout space when it exits.
Show the message queue only while a delivery is starting or running, so the first message does not flash in Queue.

Aug 22, 2026

Yaji 0.1.18

New

Add paginated conversation loading, global message search, direct-conversation history, and persisted read state.
Add Markdown rendering with code blocks, tables, task lists, links, images, and attachment references.
Add privacy-safe desktop and landing-page analytics with test coverage.
Add a central Content Security Policy for the Electron renderer.

Improved

Rework chat rendering and virtualization for smoother streaming, stable bottom following, and large histories.
Show one stable animated agent avatar and status label for each active response, including reduced-motion support.
Improve the development landing preview, conversation stories, and seeded demo content.
Add a staged hero entrance and a skeleton-to-preview reveal on the public landing page.
Extend local and remote team chat contracts for message history, search, reactions, files, and read state.

Fixed

Keep queued messages in their panel until work starts and display each user message before its matching response.
Animate queue entry removal and panel resizing without abrupt chat movement.
Keep the activity avatar visible until response streaming ends, then close it with a soft transition.
Smooth message height changes and preserve bottom scroll while streamed content grows.

Aug 22, 2026

Yaji 0.1.17

Fixed

Authenticate pinned runtime release lookups in GitHub Actions to avoid unauthenticated API rate-limit failures.

Aug 22, 2026

Yaji 0.1.16

Improved

Open an agent chat directly from incoming and outgoing exchange markers instead of showing a separate exchange history dialog.
Use Luna with low reasoning effort for every deterministic development seed agent.

Fixed

Run development-state reset tests in the Node environment so CI and signed release builds can load node:sqlite.
Keep message links, inline citations, and source references routed to the system browser, with a clear error when opening fails.

Aug 22, 2026

Yaji 0.1.15

New

Add conversational bot profile updates for name, title, and description, plus profile-based agent discovery and message routing.
Add a guided first-run onboarding flow and deterministic development seed data for integrated UI testing.
Add managed transfer manifests, shared-file references, ownership metadata, and integrity checks for message and generated attachments.
Add the shared Kobalte Select component with Storybook coverage and use it for reasoning controls.

Improved

Replace bot profile role with title in storage, local and remote Team APIs, renderer search, and agent instructions. Team permission roles stay unchanged.
Start the Auth API with the development app and select available local ports when defaults are busy.
Compact stored conversation and mailbox event history during the schema version 4 upgrade.
Simplify queued message handling by removing the paused queue state and resume action.

Fixed

Validate copied attachment size and SHA-256 data, and remove bot-owned generated files when an agent is deleted.
Improve agent settings controls, reasoning selection, profile editing, and file reference rendering.

Aug 22, 2026

Yaji 0.1.14

New

Add full-window P2P Remote Control for active server members, with shared mouse and keyboard control, four concurrent sessions, monitor selection, hide and resume, retry, and explicit disconnect.
Bundle pinned Sunshine and Moonlight Web runtimes built from source, with immutable artifacts, corresponding GPL source, checksums, SBOMs, and build provenance.
Add speech-to-text message input with local Whisper model preparation.
Add universal server invitation links and updated account, server, queue, attachment, and conversation controls.

Improved

Changed the project license from Apache-2.0 to PolyForm Noncommercial 1.0.0.
Publish the OpenBot application for macOS only in this release while Windows application packaging is paused.
Start Remote Control only from the server header and keep a hidden session active until the user disconnects or changes servers.

Removed

Remove QuickDesk, VNC, noVNC, remote passwords, and view-only remote access paths.

Security

Authorize Remote Control through active team membership and one-time in-memory viewer grants.
Verify the local Sunshine TLS chain and pin the exact generated certificate.

Aug 16, 2026

Yaji 0.1.11

New

Render Markdown and plain web links with site favicons, safe fallbacks, and system-browser opening.

Fixed

Automatically unarchive stored Codex sessions before resuming work or reading conversation history.
Use trusted Chromium input events and one consistent page and network identity in the embedded browser so X sign-in and account confirmation work.
Send signed-out X landing pages to the stable login route while preserving signed-in sessions.

Aug 14, 2026

Yaji 0.1.10

Fixed

Detect Codex and Claude in current Windows installer locations and in npm paths that contain spaces.
Report a CLI that exists but cannot start instead of incorrectly reporting it as not installed.

Aug 14, 2026

Yaji 0.1.9

Fixed

Allow Codex or Claude selection on the initial setup screen while provider checks run or setup is still required.

Aug 14, 2026

Yaji 0.1.8

New

Add GitHub-hosted Windows x64 CI builds and unsigned NSIS release installers.
Add Windows package launch, metadata, updater, and Electron fuse checks.

Improved

Publish macOS and Windows assets together only after both release jobs pass.
Detect local Codex and Claude CLI installations and enable installed updates on Windows.
Keep native window controls visible on Windows.

Aug 14, 2026

Yaji 0.1.7

Improved

Make the message composer grow up to six lines and preserve multiline typing and paste input.

Fixed

Prevent horizontal scrolling in chats and wrap long URLs and paths inside message bubbles.

Aug 14, 2026

Yaji 0.1.6

Fixed

Keep the first sent message visible and close onboarding after a successful send.

Aug 14, 2026

Yaji 0.1.5

Improved

Require an explicit model choice before a new empty agent can accept messages.
Show the specialty step immediately after model selection and after creating an agent.
Allow browser and settings panels to expand while keeping a usable conversation area.
Close an agent settings panel when switching chats and use clearer provider marks in model pickers.

Fixed

Show complete Claude responses immediately when stream deltas are missing or incomplete instead of requiring a chat refresh.

Aug 13, 2026

Yaji 0.1.4

Fixed

Present embedded browser requests as standard Chrome requests so X login and signup flows work.
Restart expired X onboarding routes from the stable login entry after an app restart.
Persist embedded browser tabs, active tab selection, URLs, and agent ownership across app restarts.
Revalidate top-level browser navigation to avoid stale cached pages.

Aug 13, 2026

Yaji 0.1.3

New

Agent headers, settings, and onboarding now use one model picker with provider availability, CLI version, and account details.

Improved

Changing the preferred provider now updates the active account details and default model immediately.
Development mode now watches source files for changes.

Aug 13, 2026

Yaji 0.1.2

Improved

New Claude agents now use Claude Opus 5 as their default model.
Agent onboarding and runtime settings now use one consistent compact card layout.

Fixed

Creating an agent now opens its settings panel immediately.

Aug 13, 2026

Yaji 0.1.1

New

Claude CLI support with automatic Codex and Claude availability checks.
Per-agent provider and model selection during onboarding and in agent settings.
First-launch provider selection with macOS permission status and later account-menu access.

Improved

Simplified provider selection to show clear availability without decorative provider cards.
Development state reset now deletes OpenBot Dev state without creating a backup.

Fixed

Shutdown now waits for active queue writes before closing local storage.

Aug 13, 2026

Yaji 0.1.0

New

Signed GitHub Releases update pipeline with in-app availability, download progress, and restart-to-install controls.
Public repository documentation, community health files, CI, and draft release automation.
Apache-2.0 licensing and an attribution notice for Norbert Bodziony.
Local Codex App Server lifecycle and ChatGPT subscription authentication.
Per-agent context-budget monitoring and proactive App Server thread compaction.
Persistent agents with independent threads, workspaces, profiles, models, and reasoning settings.
FIFO queues, agent-to-agent messaging, replies, reactions, attachments, and file transfers.
Embedded browser control and optional macOS Computer Use integration.
SolidJS desktop interface with resizable agent, browser, and settings panels.
Explicit first-launch consent before the full-access Codex service can start.
Local ZIP backups, privacy-safe diagnostics, release SBOMs, and multi-version macOS CI checks.

Improved

Standardized the product name and all user-facing branding as OpenBot.
Replaced the remaining third-party-inspired avatar SVG with an original OpenBot placeholder mark.

Meet your first teammate

Run Codex, Claude, Gemini, Grok or your own model with the plans you already pay for, each with its own workspace and context. Your work stays on your computer.

Download Yaji Contact

What's new in Yaji

New1

Improved4

Removed1

Fixed107

Security1

New19

Improved19

Fixed108

New22

Improved37

Fixed89

New7

Improved3

Fixed89

Fixed4

New45

Improved17

Fixed130

New23

Improved36

Fixed67

New12

Improved51

Fixed42

New34

Improved36

Fixed57

New6

Improved15

Fixed44

New16

Improved18

Fixed47

New29

Improved24

Fixed66

New33

Improved26

Fixed43

New18

Improved22

Fixed37

New7

Improved10

Removed1

Fixed12

New17

Improved21

Fixed19

Security1

New6

Improved15

Fixed22

New18

Improved17

Fixed77

New6

Improved13

Fixed19

New4

Improved1

Fixed1

Improved1

Fixed11

New7

Improved9

Fixed38

New5

Improved1

Fixed17

New15

Improved36

Fixed31

Fixed5

New15

Improved4

Fixed41

New32

Improved15

New

Improved

Removed

Fixed

Security

New

Improved

Fixed

New

Improved

Fixed

New

Improved

Fixed

Fixed

New

Improved

Fixed

New

Improved

Fixed

New

Improved

Fixed

New

Improved

Fixed

New

Improved

Fixed

New

Improved

Fixed

New

Improved

Fixed

New

Improved

Fixed

New

Improved

Fixed

New

Improved

Removed

Fixed

New

Improved

Fixed

Security

New

Improved

Fixed

New

Improved

Fixed

New

Improved

Fixed

New

Improved

Fixed

Improved

Fixed

New

Improved

Fixed

New

Improved

Fixed

New

Improved

Fixed

Fixed

New

Improved

Fixed

New

Improved

Fixed