Release notes
Every OtoDock release, newest first, as its changelog records it (the changelog of OtoDock 2.0.0). Before you upgrade, read the entries that change how a running install behaves, the Upgrade notes of each release you cross, then follow Upgrading.
2.0.0
Released on 2026-11-21. The release on GitHub.
Upgrade notes
- Before you update. Take a backup with
backup.sh. Update when nobody works: running chats and tasks end. The update saves the database first. If it fails, it puts the old release back or prints how to go back. An install older than 1.9.1 first runsbash upgrade.sh --to 1.9.1. - Docker installs. Chats, tasks and tools now run in a new container,
otodock-compute. The update copies your tools into it, so the first start takes a few minutes. The new images need about 11 GB of free disk. - Docker installs with their own settings. If
.envsetsOTODOCK_PROXY_MEM_LIMIT, that cap no longer covers chats and tools: setOTODOCK_COMPUTE_MEM_LIMITfor them. Ifdocker-compose.override.ymladds folders tootodock-proxy, add them tootodock-computetoo. - Installs from source. Stop OtoDock, run
scripts/dev-setup.sh --service, then start it again. Without it no chat runs. - Paired computers update themselves and restart once. With automatic updates off, press Update now under Admin, Remote Machines. Keep Windows laptops plugged in. A Windows computer with Python older than 3.11.4 (
python --version) needs Python 3.12 and a new pairing. - SSH keys. An agent's SSH key opens only the hosts an admin gave it. It no longer signs git commits.
- New models. GPT-6.2 Sol and Astra replace GPT-6.1 Sol and GPT-6 Astra. Claude Fable 5.2 replaces 5.1. Agents, chats and tasks move by themselves. If you turned an old GPT model off, turn the new one off too.
- Video and pictures. Video now defaults to Gemini Omni Flash. Veo needs a fal.ai key on Video Generation, under Admin, MCP Servers. Pictures now count toward the limits under Admin, Usage.
- Agents with a phone line. Callers get only what a manager ticks on the External identities card, in Agent Settings, Configuration. Until then they get the call tools and their notes. A notification names these agents. Each call now ends with a model summary, which the card turns off.
- Apps shared by link. A link to an app that reads the agent's files stops opening. After the update a notification names these apps and says how to open them again. If a button stops working for link visitors, ask the agent to update the app.
- Personal only agents can no longer have accounts of their own, such as a Google account connected for the agent. The update removes them. People's own accounts stay. To keep them, set the agent's visibility to Personal + shared in Agent Settings, Configuration before you update.
- Old uploads are deleted. Files uploaded into chats and untouched for 180 days go to the Recover bin for 7 days, then are deleted. This starts 30 days after the update. Change the days or turn it off under Admin, Setup, System Settings.
- Delegated work. A result that comes back to a chat in Ask mode waits for your approval. Chats an agent starts for you use your own AI account.
- Check your settings. In 1.9.1 a field could save its default by mistake. Under Admin, Setup, check that General, Security, System Settings and Phone show the values you set.
- Tools in their own container, such as File Tools, stop after 15 idle minutes and start again when used.
- Browser (Camoufox) stops reaching your local network once it updates. Its Internal network access switch, on its card under Admin, MCP Servers, allows it again.
- Rare setups. A login gateway in front of OtoDock must let
/v1/api/through. Behind a SOCKS proxy, set an http proxy inHTTPS_PROXY, or tools that need a key are off. Symbolic links to folders inside an agent's folder are ignored: move the folders back.
Added
- Outside people. On an agent's External identities card, a manager picks what a caller or a link visitor gets. The Conversations tab lists every caller.
- Chats go on after a restart. Later updates wait up to ten minutes for open work.
- Your own API key, under User Settings, Integrations, API Keys. Another agent or a script can use OtoDock with it.
- Compute machines on other servers. Admin, Remote Machines, Add a compute machine gives the command to run there.
- Several accounts of one tool for an agent, for example a work and a private mailbox.
- Community agents and departments share one list.
- Pin a chat from your chat history.
- Apps play the workspace's video, audio and pictures.
- SSH hosts now work in chats on your own paired computer, Windows too.
- People with no AI account can open agents, apps and files.
- Delete old history, a new setting per agent. Off by default, in Agent Settings, Configuration.
Changed
- Computer control is faster. The agent can do several actions in one go.
- AI accounts. A chat picks an AI account by its model. It moves to another only near a limit.
- Claude Code 2.1.325 and Codex 0.170.0. In Ask, a Codex command that needs the internet asks first. In a chat's terminal, Shift+Tab only makes Claude Code ask for approval more often. The mode menu makes it ask less.
- A small server runs fewer chats. A chat starts only while enough memory is free.
- Calls hear first words and interruptions better. A call on a person's line shows in that person's chats.
- A silent paired computer shows as offline within 90 seconds. On Linux, a command that uses too much memory is stopped.
Fixed
- Your files. The Recover your files notice appears only for a file changed in two places at once. Deleted files stay deleted. A failed upload no longer loses your change.
- Reminders an agent set for itself arrive on time.
- Failed steps. When a command, a file read or a tool fails in a Claude Code chat, the chat shows its error.
- Terminal chats on a paired computer keep their answers in the history when the connection drops for a moment.
- Long tool jobs. A long document job or video render in a Claude Code chat no longer stops after five quiet minutes.
- Tools on a remote machine now show in User Settings, Integrations, where you add your account.
- Codex now signs in with an OpenAI API key.
- GPT Image works again.
Security
- Keys stay on the server. Chats no longer hold your tools' keys or SSH keys, except picture search's SerpAPI key. Google sign-ins stay on the server too.
- Callers get no shell and no shared workspace.
- A phone line tied to an admin. A call on it now acts with the manager role of its agent everywhere, never as an admin.
- Role changes. An open chat keeps the role its person had when it started. After you raise a role, that person starts a new chat.
- Agent knowledge. An editor can no longer write into a manager's running task, which has the right to change the agent's knowledge.
- Departments. Only an admin, the department's creator or someone who manages all its agents can add an agent to it.
- Chat modes. An agent cannot make its terminal ask you less. A message sent right after switching chats lands in the chat you see.
- Server files. Installing a tool cannot read the server's secrets, and memory history cannot run commands. After the update, look for
set asideindocker compose logs otodock-proxy. If a line shows it, read the advisories on the Security tab of OtoDock/oto-dock on GitHub. - Files others wrote. A file or shortcut placed by another person, a session or a paired computer can no longer make the server read or change a file outside its folder. It cannot stop the sync either.
- App buttons. Typed text acts with the rights of the person who typed it, never the app's.
- Direct LLM chats reach only the folders they were given.
- GitHub. For an agent limited to chosen repositories, its git and GitHub commands are checked against that list.
- Phone call logs keep no words of a call. On Docker the update clears the old ones. From source,
phone/phone-server.logkeeps them until it rotates. - Open documents close or turn read only within seconds when a person loses access.
- Security updates: PostgreSQL 16.16 and Node 26.13.1, on a newer Debian 13 base. The database keeps its data.
1.9.1
Released on 2026-11-04. The release on GitHub.
Upgrade notes
- Memory topics over 32 KB stop growing until the agent shrinks them, and one edit adds 2 KB at most. Admins set both limits under System Settings.
- Only editors and managers now see an agent's own tasks, triggers and runs, in apps too. Others see their own. A shared task or trigger picks who is told of its runs, under Who is told: a new one tells editors and managers, existing ones everyone.
- Scheduled tasks and delegated lanes can now create images, music and video when the agent has those tools.
- A chat you delegate can run on your own computer when you ask for it ("do this on my laptop"). There it runs as you, with your AI account, browser and connected accounts. While that computer is offline the lane says so. Only you can continue it there.
- Every agent now allows every AI engine, each person on their own account. A limit you set before is lifted once: untick engines in Agent Settings, Configuration to set it again. The agent's own tasks still need the platform's subscription.
- A Mac paired before 1.9.1 starts chats again after the update. Pair it again so every tool installs: its owner removes it in Remote machines, pairs it and gives back its agents, browser and computer control. macOS asks again for its permissions.
- Paired computers update themselves to 0.5.168 at reconnect, one restart each. Browser and computer control install again on their own.
- Dictation in Safari, on an iPhone or iPad and with the Automatic language now bills your speech provider. To stay on the device, pick a language in the microphone menu, and in Safari or on an iPhone or iPad set Dictation to Native under User Settings, Audio.
- Restored an older backup? If Slack, Notion, Zoom or GitHub events stop, ask OtoDock support to reset them.
- GPT-6.1 Terra and Qwen 3.8 replace GPT-5.6 Terra and Qwen 3.6, and GPT-6.1 Sol joins Direct LLM. Agents, chats and tasks on them move at the server's first start. Terra costs less, Qwen 3.8 a third more. Add an old model back as a custom model to keep it.
- Sign-in addresses (Setup, Security, OAuth Bearer Allowlist) are no longer added by OtoDock itself. Each MCP adds its own and takes over a matching row an admin had, one installed before 1.9.1 at its next update or off and on. Remove custom addresses there drops the admin rows.
- The server's images run on Debian 13. An SSH host with a DSA key stops working: give it an ed25519 key. An agent's Python 3.11 environment stops starting: delete it, the agent makes it again. An MCP needing Python 3.11: remove and install it again under Admin, MCP Servers.
- On a source checkout the document editor now answers on its own computer only. A reverse proxy on another computer
that reaches port 9981 needs
COLLABORA_BIND_IP=0.0.0.0in config.env before you update. The first start then rebuilds the file tools: twenty minutes with 6 GB free, while the rest works.
Added
- A department installs from the community in one step. Its agents, shared library and paused schedules come too.
- An agent can propose a community agent, a department, or moving an agent into one. The Agent Creator tool searches the community catalog first and shows a card in the chat. Nothing changes until you press it.
- Ask an agent where a setting is, and change it in the chat. Its card lets each person in the chat change theirs. The agent never sees the value. A manager turns the Settings tool on under Agent Settings, MCPs.
- An agent can offer you a choice as a card with buttons. A button answers, or does one thing with your own permissions, and nothing happens until you press it.
- The iOS app, on TestFlight. One app for more than one OtoDock, your company sign-in, dictation, notifications and alarms with sound. Both apps ask for a rating, twice at most.
- Documents agents make are checked and look closer to the preview. Word gains tracked changes and comments, slides gain themes, spreadsheet formulas get their values, and PDF forms fill in.
- An admin can let every role's chats run on a platform machine.
- Cartesia voices in the Text-to-Speech tools, at a chosen speed and, in English, emotion. The microphone menu says which recognizer listens.
Changed
- The Browse Community Agents page shows the apps a template brings. An app copied from another agent follows its source.
- An editor or a manager can share one of their own personal memory topics with everyone on the agent.
- Your own subscription now works on every agent that allows the engine. Each agent lists first the engines you use most there. A new chat opens on the first. In Configuration, Only allow local models keeps an agent local.
- Platform only joins the chat audio policy, an admin choice that sends everyone's voice and dictation through your speech provider.
- Greek is heard better in a spoken conversation, and its speech shows on the Usage pages under the person who spoke.
- Delegated work shows as a card in your chat, and each delegated chat gets a short title with an icon. A lane from your chat has your chat's tools and settings.
- The Notebook replaces library bulletins (the old files stay). Agents one delegation step apart share dated notes there.
- Terminal chat windows fold into icons you can reorder, and each chat keeps them. The company map opens on the agent you came from.
- Claude Code 2.1.312 and Codex 0.166.0, installed on every machine at reconnect. Claude shows its plan review in Plan.
- Reports of replies (Admin, Reports). Admins mark one handled or delete it. It is deleted a year after it was filed.
Fixed
- Dictation on the device keeps every phrase and survives a pause, and says so when it hears nothing.
- Photos and files sent with a chat's first message stay in the chat, reach an agent in a terminal, and are not sent again on Resume. In a terminal chat a message shows once after a reload.
- Background work keeps its badge until it ends and asks for approval at once. A follow-up reaches a running lane.
- A Mac pairs from a fresh install, and computer control asks for its two macOS permissions. The browser starts on a Mac where Chrome was never opened.
- Paired computers. One that showed online while refusing every tool call now repairs its link. A chat shows only its computer's install, a first workspace copy counts its files, and a message to a silent one ends in a card.
- Direct LLM, Docker MCPs and the Local Browser. A chat stops waiting for a tool that never answers. A Docker MCP's update shows its progress and runs once. The Local Browser no longer stops: a stuck tab restarts it.
- A self-hosted Postiz signs in. Removing OtoDock at a vendor disconnects the account and stops its events. Accounts a vendor stops accepting say Reconnect needed. Admin, Users lists last sign-ins. A Claude login ending soon offers Reconnect.
- Older chats and task runs load as you scroll the sidebar, and a search result opens at the match.
- Small display fixes. A check's condition and a weekday schedule read as sentences. Choosing where a chat runs clears a phone's message box. Mouse-only buttons show on tablets and foldables. The offline computer notice can be closed.
- Codex plans in Plan again. A question from an agent ends its reply and asks you. Cancel on a plan stops the agent.
- Phone calls. An agent can call through another agent's line, and that agent speaks on the call. Calls an agent placed can be transferred too, and a call's result reaches the chat at once.
- The Android app. From Android 13, turning notifications off stops alarms too: keep them on for your alarms. Videos turn with the phone. You can switch servers while a reply is still coming.
- Ubuntu 26.04. Agents run again on a server or a paired computer installed on it.
Security
- A document cannot make the server fetch what it only links to. pyasn1 0.6.5 closes four advisories.
1.9.0
Released on 2026-10-25. The release on GitHub.
Upgrade notes
- Accept Edits is now Auto, and Default is now Ask. Ask works as Default did. In Auto the engine reviews all but
ordinary work. New chats start in Auto or your last pick for the agent, and
otodockon a paired computer in Auto. - Paired computers get OtoDock 0.5.158 and Codex 0.163.0 at reconnect. Until then, moved chats keep only a summary. A Windows one on Python 3.10 loses its tray icon and its Pause. Install Python 3.11 or later and pair it again.
- Editors never run a chat on an admin machine. Their message moves it to the OtoDock platform or their computer.
- Blender's code tool no longer asks on every call. It runs unasked where a computer's App connectors grant is on.
- Automatic language detection is on for dictation set to Automatic and for routes made since 1.8.0. Older routes need it turned on in Setup, Phone. While it decides, for up to 20 seconds, each dictation or call runs one or two extra speech streams. Your provider bills them.
- Upgraded installs keep a 60 or 180 second dictation limit. The extra streams count toward it. New ones get 300.
- Chats idle for 180 days lose their working files on paired computers too, and continue from their history. To keep the files, turn this off before upgrading in Setup, System Settings, Storage and Retention.
- Outside services need public DNS answers. Private or fake ones stop the catalogs, push notifications, usage
checks, Microsoft sign-in and Twilio. Set
HTTPS_PROXYin.env, orconfig.envfrom source, to a proxy that looks names up itself. - Four new optional settings in
config.env.example:HTTPS_PROXY,NO_PROXY,COLLABORA_SSL_TERMINATIONandWOPI_ACCEPT_UNSIGNED. The last keeps your own Collabora working until it has its key. DeleteCOLLABORA_FRAME_ANCESTORS: nothing reads it now. - The document editor moves to Collabora 26.04, which needs Docker Engine 26 and Compose 2.26 or later.
upgrade.shchecks both and stops before any change. By hand, follow the upgrading guide, and on https setCOLLABORA_SSL_TERMINATION=truein.envfirst. - A
docker-compose.override.ymlthat pins an older Collabora, or replaces itsentrypointorextra_params, breaks the editor. Remove those lines first. - Node 26 and pnpm 12 come with the upgrade. Compiled Node modules in an agent's project need
npm rebuildonce. - OtoDock refuses to start when its agents' sandbox cannot run. It needs
/dev/net/tun, which Docker installs have unless you changed the proxy'sdevicesorsecurity_optlines. On a native install, runsudo modprobe tunand listtunin a file under/etc/modules-load.d/. - A Meeting Rooms limit of 0 now turns meetings off. It used to mean four. Clear the field to keep four.
- Apps built on OtoDock's styles take the new look. If one looks wrong, ask the agent that built it.
- Going back to 1.8 signs everyone out, and two-factor is asked again.
Added
- The Android app comes to Google Play after this release. Notifications and alarms reach your phone, and agents can ask where you are once you allow it.
- Choose where a chat runs: the OtoDock platform or your paired computers, and for admins an admin machine. The icon beside the model shows it. Move a chat later, or ask the agent to.
- Tasks and delegated work on your paired computers. When that computer is offline, the run goes where the agent usually runs and says so.
- Plans on Codex. A Codex plan ends with buttons to implement it in Auto or Ask, edit it or reject it.
- GitHub Apps. Use your admin's GitHub App, or OtoDock's with hosted sign-in. A manager picks what the agent's own GitHub account reaches, or makes it read only. Each person picks for their own chats. Needs GitHub MCP 2.1.0.
- An admin command line for the day nobody can sign in.
otodock-admin.shresets a password, turns off a lost authenticator and prints support information. On Docker, put your release's copy from GitHub in the install folder. - Report a reply. A flag under each agent reply sends it to Admin, Reports, never to OtoDock.
Changed
- A new look. A new logo, font and dark theme, with clearer colours and plainer words. Settings share one layout, with long explanations behind a small info button. A setting that saves at once shows Saved beside it.
- Tools follow their own labels. One marked as only reading runs without asking, as does a tool a computer's owner granted. Ask asks for the others. Auto runs one whose changes can be undone and reviews the rest. Six OtoDock tools, like changing a department, ask in every mode.
- Shared chats on your own computer. On a Shared only agent, a member can move a shared chat to their computer. Others can only read it there. The agent's accounts are never copied to it.
- Language detection follows the speech. Where the speech service scores its words, the clearer words decide, so an accent no longer switches the language. List only the languages your callers speak, under Setup, Phone.
- SSH hosts follow the agent's roles. On the server, its managers and editors can use the hosts an admin added.
- Apps your agents build match OtoDock's look, or your client's brand when you ask.
- Every engine shows its progress. Codex sums up its thinking, and the chat says Working while nothing new shows.
- Dialogs work with the keyboard, and every refused or failed action says why.
- If an upgrade fails before the new release starts,
upgrade.shputs the old release back and says what failed.
Fixed
- Resume. When a reply ends early, Resume asks the agent to carry on instead of sending your message again.
- Chats. A reply no longer splits around background work or a meeting's start. A question answered in one tab closes in the others. A long reply survives a restart, and a chat survives a slow or dropped computer.
- Phone calls and dictation. Calls catch Greek words they used to miss. Long dictation keeps every word, and a slow connection says so. A call ends with a short goodbye.
- Also fixed: a save never overwrites a file a paired computer is still writing. It goes to the Recover bin. A crashed tool container shows as stopped. The terminal no longer repeats autocorrected text.
Security
- A person's Codex task, meeting or call acts as that person. It used to act as the agent.
- Agents cannot wipe the root or a home folder, however it is written. rsync to other computers works again.
- The bundled editor opens a document only for the person signed in. A link from its log opens nothing.
- Calls to outside services go only to the address OtoDock checked, even if DNS changes.
- MCP servers can no longer fetch your GitHub token. Only git and gh in the agent's shell can.
- A Codex chat's first command runs inside Codex's own sandbox again. From 1.7.1 to 1.8.1 only OtoDock's held it.
- A sign-in link or token made for one purpose is refused elsewhere. Older ones work until they expire.
- Tasks and triggers that run as the agent reach only its shared files. One that wrote a person's files must run as that person. One that changed the agent's settings needs a creator who manages it.
- Direct LLM chats offer only the tools the person may use.
- A failing app shows its visitors a short error, never its code.
- The libraries behind images, certificates and web requests are updated to close published vulnerabilities.
1.8.1
Released on 2026-10-14. The release on GitHub.
Upgrade notes
- None.
Changed
- The box that opens after an update names each new feature in a line or two.
- The README's first picture is captioned "Full office suite, plus photo and video editing."
1.8.0
Released on 2026-10-13. The release on GitHub.
Upgrade notes
- Claude Code 2.1.296, Codex 0.162.1 and OtoDock 0.5.147 on paired machines. Paired machines update when they reconnect. Turns running on them end. Agents pinned to Haiku 4.5 move to Haiku 5.5 at boot.
- Cartesia speech uses the sonic-3.6 model. A retired model id moves to it. Both speech providers now voice every language. Your voice picks stay.
- Connecting an account needs
DASHBOARD_PUBLIC_URL. Set it to the address people use for the dashboard. A connect started from another address is refused and tells you the address to use. - Catalog MCP updates wait seven days after a release. The Update button still applies at once. An MCP's default settings apply until you save your own value.
- An older server refuses a database a newer release already upgraded.
OTODOCK_ALLOW_NEWER_DATABASEoverrides that. A yes or no setting in the config accepts 1, yes or on. Any other word stops the start. - Log lines from the server are named
otodock, no longerclaude-proxy.MAX_PARALLEL_SPAWNSset to 0 on a delegation MCP now allows no workers. Clear it to get the default of four.
Added
- An agent can end a call or transfer it to a colleague. On a call another agent started, it can also ask that agent. The default call prompts mention the new tools. A prompt you never edited takes the new text.
- Calls can follow the caller's language. A route can identify it from the first words and switch the voice. Rare route fields sit under Advanced. A route can pin one voice. Dictation gains an Automatic language.
- A call placed from your chat runs inside it. On a route tied to you, the person's words arrive as messages and the agent speaks back. The chat resumes when the call ends.
- Pictures and files an MCP returns reach the chat. They are saved in your workspace and shown with the result. MCP results show their text on Claude again.
- Back up and restore the whole install by script. One dated folder holds the database, the settings, the agents, the SSH keys and the accounts. A restore refuses a newer backup. The installer checks the free disk before it pulls images.
- Admins see the version and new releases. A daily check names a newer release. A
dialog shows what is new after an update.
OTODOCK_RELEASE_CHECKturns the check off. - Folders upload whole, dropped or picked, in the workspace and in the chat.
- An outbound call can ask for a PIN. The person who answers enters it before the agent speaks. Wrong codes end the call. A switch saves a call's outgoing audio, ten files kept, never the caller's voice.
- Claude Haiku 5.5 on Claude Code and Direct LLM. A 1M context and adaptive thinking at a twentieth of Sonnet's price. GPT-5.6 Terra and GPT-6 Luna get a 1.05M window on Direct LLM. Prompts past 272K tokens cost OpenAI's long-context rate.
Fixed
- Paired machines sync safely again. Windows machines send files back, broken since 1.7.1. Files the bug dropped return once the machine updates. A file a machine cannot read never counts as deleted.
- Calls repeat nothing and lose nothing. A first call keeps its start. A transfer survives the hangup. Greek dictation shows at once. An unserved call hears an apology. The log says who hung up.
- Messages you send are kept, in order. Messages typed while a chat gets ready go out after its first message. Voice, phone and meeting turns wait their turn.
- Delegated workers report when done. A run waits for its background work. A stuck permission ends in ten minutes. A Stop ends the run at once and reports it as stopped.
- The dashboard is steadier. Fields keep their size on phones. The map shows only what you may see. A document saves at once and reloads after idling. A waiting permission card pings you. The file browser shows synced and uploaded files.
- Sessions follow what you set up. Claude and Codex read each repository's instruction file. A retired model runs on its successor. Task runs see the time.
- Also fixed: a paste from Word, Excel or LibreOffice keeps its text. Sonnet 5.5 cache reads cost $0.10. The recover bin survives an upgrade.
Security
- GitHub tokens leave the agent's environment.
gitandghfetch the account's token per command. A disconnected account stops working on paired machines too. GHSA-9m79-q4pv-8hfj, thanks to @jourasharsy. - A file from a paired machine lands only where the person may write. A document opened from a person's own machine is edited by that person only. GHSA-6vfw-92r2-jw83.
- An agent cannot change the rules that judge its tools. The command
rm -rf /is refused on every engine and in every mode. A permission check that cannot run denies the call. - Release images carry a signed provenance and an SBOM. Each image is booted before
its tag is written.
gh attestation verifychecks them. SECURITY.md explains how. - Sign-ins are single-use. A 2FA code works once, across restarts. A double sign-in opens one session. An account connect finishes in the browser where it started.
- Web requests the server makes go to the address it checked. Webhooks, app traffic, event lookups and push messages resolve a host once and only to public addresses.
- Document tokens are narrower. A token opens only its own document. The editor refuses a sign-in that was signed out everywhere or is due a new password.
- The file tools take cairosvg 2.9.1 (CVE-2026-107378).
1.7.1
Released on 2026-10-07. The release on GitHub.
Upgrade notes
- Enter adds a new line in the chat box. Shift+Enter, Ctrl+Enter or the Send button sends the message. On a phone or tablet the Send button sends.
- App shares made before 1.7.1 wait in "Shared with you". Accept them once more.
- Paired machines update their OtoDock software during the upgrade. Turns running on them end. The document editor's container is recreated once.
- A new
JWT_SECRETunder 32 characters is refused. One already in use only warns.CREDENTIAL_ENCRYPTION_KEYnow works on Docker installs too. Old secrets still open. - Behind a reverse proxy, a banner tells you the
TRUSTED_PROXYvalue to set. The installer tells you too.
Added
- Share an app with a person, an agent or a department. You pick their role. They accept it from "Shared with you" in the notifications panel. Agents can use the apps placed in them.
- Documents open in a pane beside the chat. One tab per document. Every version the agent delivered in that chat is there. The pane opens by itself when a turn delivers a document.
- MCP servers hosted by a vendor sign in with OAuth. You sign in once in the dashboard. The token is added on the way to the vendor. It is never written to a session's files.
- An MCP can move to a new catalog source. Admins accept the move on the MCP Servers page. Settings, credentials and assignments are kept.
- A delegated worker hands files back to the workspace of the agent that asked. If the worker stops to ask a question in a chat nobody is watching, you get a notification.
Changed
- Claude Code 2.1.289, Codex 0.160.0 and OtoDock 0.5.138 on paired machines. Paired machines update when they reconnect. A background command in a scheduled task stops after 30 minutes.
- Sonnet 5.5 and GPT-6.1 Sol replace Sonnet 5 and GPT-6 Sol. Agents pinned to the old ones move over at boot. Chat titles on Anthropic use Sonnet 5.5.
- When an agent asks you a question, it waits up to a day for your answer. A server restart under two minutes does not break a running tool call.
- Chats stay fast with a long history, many open tabs or a burst of messages.
- SSH tools need the editor role or above. Existing assignments are kept and marked.
- Shared apps: you can remove a share you accepted. Placed apps have a teal border. Admins see every share on the Shares page.
- Webhooks from GitHub and API-key senders may be up to 25 MB.
- The Local Browser can open local services you add to its site list. An MCP install that hangs is rolled back.
Fixed
- Documents keep your edits. A save is no longer lost when the editor's background save hangs. Closing the editor on a phone keeps the save too. A document on a remote machine is saved 10 seconds after your last keystroke.
- Messages typed while the agent works wait in the chat. Every tab and teammate sees them. They go out in order after a switch, a restart or a reconnect.
- A turn that cannot finish ends with a card that says why. A decline, the usage limit, an engine error, 10 minutes of silence or a lost machine. Where it helps, a Send again button is there.
- Paired machines on a slow or flaky connection keep large transfers and running turns. After a server restart, a paired machine's chat keeps its session.
- The Android app opens servers with an underscore, IPv6 or an international name. It stays put during sign-in.
- Smaller fixes cover meetings, refused chats, HTML conversion, big uploads, restore and backups.
Security
- A file sent to another agent can no longer write outside that agent's workspace through a planted link. GHSA-jfp5-m3vx-9272, thanks to @zaara2004.
- Single sign-on checks the identity token. It refuses an unverified email that another account already uses. Sign-in attempts are limited per address.
- A sign-in ends when the password changes or the role is lowered. Tokens no longer appear in page URLs or logs.
- Text that arrives from webhooks, phone callers and visitors is handed to the agent as data, never as instructions.
- A page on another website cannot make changes in your OtoDock while you are signed in.
- New phone PINs have 6 digits. Lockouts survive a restart. Plain http from the internet to the proxy or the phone service is refused.
- PyJWT 2.15.1, urllib3 2.8.0, multidict 6.9.1, Capacitor 8.5.2 and pnpm 11.28.2 fix advisories. So do Python 3.13.16, Node 24.21.0 and Postgres 16.15.
1.7.0
Released on 2026-09-29. The release on GitHub.
Added
- Apps are now a platform of their own. Every app has its own page, can have a server and a database everyone works on live, wake on a schedule, a webhook or a finished task, and talk to outside services with keys the agent never sees.
- The agent checks and deploys its own apps. The platform refuses what the sandbox would block, viewers keep the last copy until the new one answers, a change in what an app may do waits for your approval, said in plain words, and you can roll back.
- Share an app or a chat with a teammate, or with anyone by a password link. A shared chat is read-only, a link's buttons stay off until you turn them on, and a link can be set never to expire while no admin cap on share expiry is in force.
- Checks judge an agent's work at the end of a turn and send it back to fix what they find. A check is a script, an app's handler, a schema or a rubric. Agent templates come with their own apps and checks and follow new versions.
- A contributor role: files without the agent's identity. Between viewer and editor: a contributor adds files to the shared workspace, and nothing else runs in the agent's name.
- Each AI engine has a default model chosen by tier. An agent without a pinned model runs its engine's strongest tier below the frontier; every model list shows its tier.
- One GitHub subscription covers a whole organization (GitHub MCP 1.2.0, needs the Organization webhooks permission), and an agent can subscribe on its own behalf.
- Hold the microphone button to pick the dictation language. Admins can remove any user's machine.
Changed
- Upgrading from 1.6. A Docker install takes new compose files: use the upgrade script,
not a version bump alone. Behind a reverse proxy set
TRUSTED_PROXYor local-only accounts are refused, the phone ports listen on the IP your PBX dials,/docsis gone,/healthshows no versions, and new caps apply. More at docs.otodock.io/administration/upgrading. - Shared-only agents: chatting takes the editor role. Their chats run as the agent itself, so a viewer or contributor there now gets a message instead of a session. Give them the editor role, or turn on personal chats for the agent.
- Departments pick a delegation mode. Off, down only (the default for a new one), down and across, or all directions, each drawn on its card, with a reach of one level or the whole department. Existing departments keep their wiring.
- Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna replace Opus 5, GPT-5.6 Sol and GPT-5.6 Luna. Cheaper, same capabilities. Pins move at the next start, and a custom model under the old id keeps its pins.
- Claude Code 2.1.281, Codex 0.156.1 and satellite 0.5.130. Installs update on restart, paired machines on reconnect (or by hand where automatic updates are off).
- A message typed while the agent works reaches it mid-turn, wherever the agent runs.
- Mini-apps are now called apps, an app's card says whose account its buttons use, and Codex chats ask before every command and edit, as Claude chats do.
- Architectural improvements under the hood. One shared vocabulary for run statuses, placements, roles and tool kinds, and one contract every AI engine follows. A paused machine shows as paused and a usage limit ends its run.
Fixed
- Chats stay in their lane. A pick right after New Chat no longer changes the previous chat, a draft or a steered message no longer lands in another chat, dictation stops on a switch, and attachments sent while the agent works reach it.
- Paired machines. Mode and model changes take effect, a Codex chat keeps its thread across a restart, a prompt no longer lands twice, and a follow-up to a worker open as a terminal lands in that terminal.
- Wakes, automations and load. A chat woken by a check-back or a delegated task's result comes back as the person who asked, and a Codex chat resumes its conversation; the automations of a removed person pass to the admin; a hundred streaming chats or a database hiccup no longer freeze everyone else.
- AI engine setup is honest. The "Auto" model names the model that runs, a misconfigured engine fails with a clear message, and custom prices are honoured.
- Tasks and triggers. A run says what fired it, cancelling closes its output at once, and a prompt edit applies from the next run. Template tasks run right after install. Webhook addresses written with a display name fire again.
- Dashboard details. A refused photo no longer leaves the chat busy, the MCPs page shows its read-only notice, a wrong passkey password shows its error, the company map's top bar is dark in light mode, schedules say 21st, and on a phone the app menu, the approval card and a user's agent roles fit the screen.
- A slow first start no longer fails
docker compose up, which gave the server a minute.
Security
- A security review of the whole platform ran before 1.7. Every file an agent writes or the platform serves stays inside that agent's folder, sign-in and sessions are stricter, a paired machine's own OtoDock folder is off limits to the sessions it runs, and an MCP update never changes where its package comes from.
- Access checks tightened after a full review of roles, apps, tasks and calls. Nobody reaches another person's chat, terminal, call or task run by its id, a session never sees the agent's own engine settings or logins, apps and shares keep to their own app and agent, and phone calls and SSH keys take the editor role.
- The perimeter holds under hostile input. Sign-in, passkeys and webhooks are rate limited per address and never stall the platform, Require 2FA holds a passkey holder to a second factor, bodies are capped, the dashboard socket refuses other origins, and every fetch to an address a person or an agent picks checks where the host really points.
- Connecting a ChatGPT or Claude account is a dashboard action, rate limited per person, each device login in its own private directory. The Android app's native features answer only the dashboard's own page, and its release build is signed.
1.6.1
Released on 2026-09-15. The release on GitHub.
Added
- The release images ship for arm64 as well as x86-64. Raspberry Pi 5, Ampere and Graviton servers and other arm64 Linux hosts pull the same tags, the phone service included. The installer refuses an unsupported CPU with a clear message. Thanks to suranyami for the pipeline and the field test (#2).
- The pool reads each account's usage from Claude and ChatGPT. New work goes to the account whose weekly window resets first, and an account out of window for the model is skipped. Both windows show on the AI Engines cards. Owners get a notification at 90% and when a weekly limit is reached.
- Caps on subscription usage. Each user can cap their own accounts, and an admin the agent pool: a share of the weekly windows or dollars, per day and per week. At the cap the work stops or continues on API keys, as you choose. Set on the Usage pages.
- Users can connect their own Anthropic or OpenAI API key in User Settings → AI Engines and cap what it spends per week or month.
- Every MCP shows its icon and the project it wraps in the community browser, the admin MCP Servers page and an agent's MCPs tab, and the lists filter by category.
- Issue forms and a pull request template on the public repositories. A bug report asks for the version, the install kind and the logs.
Changed
- A platform setting can stop OtoDock reading subscription usage from the provider (on by default). Off, the pool balances by recorded cost as before.
- The dashboard downloads less and opens faster. Its scripts and styles arrive compressed and stay cached until the next release. The first screen carries a third of the code it did.
- A docker community MCP whose catalog manifest changed under the same image tag now shows an integration update, as node and python MCPs already did. Docker MCPs installed before this release show one such update once.
- The phone service's voice detection runs on the ONNX Runtime it already ships. Same Silero model, same behaviour, one native dependency fewer. A bare-metal phone install picks it up by re-running its setup script.
Fixed
- An open dashboard reloads itself after an update. A page kept running the previous build until reloaded by hand; it now reloads once, after any sign-in, meeting or upload still in progress.
- The wake word hears the first attempt, and agent names wake as easily as "Hey OtoDock". The spotter forgot the start of a phrase about one attempt in twelve. A listener whose engine failed to load is rebuilt, and a "Record diagnostics" switch keeps the last thirty seconds on the device.
- Connecting an Anthropic Console account is refused with the reason. Such a login carries no inference scope, so every turn failed with "Not logged in". The connect form now points at an API-key credential instead (#3).
- Subscription seat counts stay honest. Abandoned spawns leaked one seat each and blocked Remove until a restart. Seats now return, and Remove judges on live sessions or can be forced (#3).
- An agent-scoped run the pool refuses fails with the reason instead of starting on a stale credential file. The task run, meeting or call says why.
- Chat fixes. An interactive terminal's questions and plan approvals park the chat as "needs your input" instead of freezing it. A first prompt is no longer typed twice into an open dialog. A chat reattaching to a running turn after a reconnect works again.
- A satellite started from inside Claude Code spawns clean CLIs, so its chats persist. Inherited session variables had made every terminal there a nested child that kept no transcript.
Security
- File tools render PDFs with WeasyPrint 70. The renderer no longer loads EPS images and keeps the original URL fetcher for stylesheets and metadata (CVE-2026-55073).
1.6.0
Released on 2026-09-09. The release on GitHub.
Added
- GPT-6 Astra on the Codex engine. OpenAI's new flagship is a predefined Codex model. "Auto" keeps choosing GPT-5.6 Sol: Astra costs 2.5 times more.
- Local models, everywhere. A Codex agent on a paired machine can run on a local endpoint; one endpoint serves both the Direct LLM API and Codex engines and can carry an API key; chats show a local model's thinking and pass effort.
- The Direct LLM API engine catches up. OpenAI models use the Responses API
(effort with tools, reasoning shown, web search with sources, nothing stored).
Agents get file tools and the sidecar MCPs. Tools and skills load on demand
(
DIRECT_LLM_TOOL_SEARCH=offrestores the old way). Needs the current relay. - Phone routes know who is calling. Each caller gets a private folder and memory (Per caller, the default), or the call runs as a chosen platform user (role capped at manager); callers are viewers with no platform or shell tools.
- Caller data retention. Callers' folders, conversations and the call log age out together (Setup → Phone → Caller Data, 90 days, "Forget all"). Callers' folders count against a new per-agent quota (1 GB).
- Browser control can use your own browser. A remote machine can opt in to "Use my own browser": the agent works in the Chrome, Edge or Brave you are signed into, through the Playwright Extension (save its token for tasks).
- The video-tools MCP, 0.4.0 to 0.4.2. Colour-managed grading (HLG and PQ
to Rec.709, LUT chains, clarity, sharpness), a colour-managed
edit_video, analign_audiotool, Rec.709 renders at 48 kHz. Update it from the catalog. - Sandbox limits and a stall watchdog. Each local sandbox's
/tmpis capped (SANDBOX_TMP_SIZE_MB, 4096 MB, at most half the host's memory). A watchdog logs the blocking stack on event-loop stalls (LOOP_WATCHDOG_THRESHOLD_S, 2 s).
Changed
- Codex CLI 0.153.4 and Claude Code 2.1.263. Paired machines update at their next reconnect. Codex's plan tool stays on so its checklist keeps working.
- Cost shown only when the turn cost money. API-key and hosted relay chats keep the badge, subscription and local-model chats hide it; usage is unchanged.
- Phone routes on upgrade. Existing routes come up in Per caller mode: a caller who worked in the shared space gets a private one and no longer sees the shared memory (tie the route to a platform user to keep the old way).
- Phone daemon (rebuild it with the proxy): call-log retention follows the
caller-data window; outbound calls wait for their pre-generated opening; the
PHONE_AGENT_ROLESsetting is gone; admin-only agents no longer run as admin. - "Primary" is gone from AI Engines. The model decides the provider. System Default prefers a provider with a subscription, so a local-only install no longer defaults to a hosted model. Stored primary flags are cleared on upgrade.
- Claude models on Direct LLM search the web with the basic search tool again.
- Satellite 0.5.118 (fleet auto-update): unattended Codex sessions on paired machines run the same permission floor as on the server; lean skill cards.
Fixed
- Dictation no longer repeats words. The 1.5.0 fix for vanishing words could print the last words of a phrase twice. It now commits exactly what was heard.
- Codex on local models. Discover finds an Ollama server's models again. On Ollama 0.34 the MCP tools load on demand (0.34.0-rc1 cannot call them yet, reported upstream). A slow first prompt no longer dies at five minutes.
- Codex sessions start clean. The first turn waits for every MCP server. A stale ChatGPT login no longer shadows an API key or a local model. llama.cpp and LM Studio are told they drop MCP tools. Meeting tools no longer prompt.
- Direct LLM chats. A chat on an agent pinned to a paired machine no longer believes it runs there. Switching provider mid-chat works. Web search, fetch and code rows finish. A cut-off tool call no longer breaks the next request.
- Phone calls. No two-second dead air after each sentence with ElevenLabs
voices (
close_at_flush). The ambience plays through gaps. A call ends at the agent's end marker. A timed-out audio read no longer leaves the call deaf. - The proxy no longer stalls on a slow database disk. Status, turn saves,
chat actions and log writes run off the event loop. Tunneled MCP streams are
no longer cut every 15 minutes.
send_filessees same-turn remote files. - Server sandbox. Bash commands that start with a variable assignment run
again. The permission gate looks past
do,then,ifand( … )wrappers. The session scratchpad under/tmpis writable. - Company map. Zooming out of a department lands on the same whole-map view that paging between departments uses, with no second zoom-out step.
- Meetings end each turn at the routing call. A participant that kept
working after
direct_tostalled the meeting, and a summary written afterend_meetingwas lost. Later tool calls are refused, the summary is kept. - Smaller fixes. New models show in the pickers at once. No "Default Session
Mode" on a Direct LLM agent.
write_xlsxcharts and conditional formatting come out as asked and survive later edits.
Security
- Sessions without a human user (tasks, triggers, meetings, phone) are tighter: permission state clears when a local session closes, agent configuration and warmup need the master key, the memory API refuses foreign-agent sessions, phone-minted tokens die at hangup, delegation refuses external principals.
- Uploads, context files, recover-bin restores and previews re-check their destination on the resolved path (a planted symlink cannot redirect a read or write); the file tools check every path with the platform, including container-absolute ones. Phone numbers no longer appear in logs.
1.5.0
Released on 2026-09-03. The release on GitHub.
Two headline additions — your agents on your own machines, and your agents on the phone — plus departments, shared knowledge libraries, a 3D company map, live voice conversations, and a long list of fixes.
Added
- Remote machines. Pair a Linux, macOS or Windows computer with a one-line install and run your agents on it, with full access to that machine's files, apps, network and tools — same chat, same live streaming, same dashboard. The connection is outbound-only (no open ports), pairing codes are single-use, and every capability is an explicit grant. Files sync both ways (generated build folders are skipped automatically), idle sessions survive platform restarts, machine cards show each machine's CLI versions, and the satellite keeps itself up to date. Two tools ride along for agents on a machine: browser control (drive a real, logged-in Chrome/Edge) and computer control (screen, mouse and keyboard).
- Phone calls. Agents answer and place real phone calls, through Twilio (paste an account SID and token; calls enter through your existing dashboard URL) or your own Asterisk/FreePBX. Calls run on the live voice pipeline: natural replies, talk-over interruptions, a spoken line before each tool call. Routes can require a keypad PIN (lockouts after repeated failures) and keep a 30-day call log. The phone service is part of the standard install and idles until telephony is configured; its image is published with every release. A phone-calls tool places outbound calls from chats and tasks, and a transcription tool turns audio and video into text and SRT subtitles.
- Live voice conversations. A phone-style mode in chat: open mic with echo cancellation, instant barge-in (even while the agent is silently using a tool), a live transcript in the composer, short spoken replies while documents and dashboards still render, click-to-edit drafts, and agents that can end the call themselves. Say "Hey OtoDock" (or "Hey" plus an agent's name) on any dashboard page to start one hands-free — detection runs entirely in your browser, per-user opt-in. Needs the phone service's voice engine; it replaces the earlier hands-free mode.
- Departments. Create departments with named levels (up to 8), give each agent a department and level, and delegation wires itself: peers to peers, each level to the one below or the whole subtree, and results back up to the level above. A chain guard stops delegation loops (depth capped at 4), the Agents page groups agents by department, and each agent's prompt states its place in the org.
- Delegation between agents. Hand work to another agent as a visible
parallel session you can watch, steer and continue; the result comes back
as a report.
send_filesdrops workspace files into another agent's inbox. A scheduled run can read what the agents on its roster did (tasks, sessions, triggers, notifications) — enough for a CEO agent's morning briefing — while your own sessions read only what you could see yourself. - Shared knowledge libraries. Share an agent's knowledge folder, or any
subfolder, under a name of its own and attach it to other agents read-only
or writable, on remote machines too. Writable copies flow edits and
deletes back to the source at the end of the turn; read-only copies are
healed, with stray edits kept in the Recover bin. A
bulletin/<library>.mdfile is loaded into every attached agent's context — a standing team briefing. Manager-created scheduled tasks and delegated workers can write knowledge as well. - A 3D company map. The Agents page is a moonlit world where each department stands on its own base: agents arranged by level, activity heat and live "responding" pulses, delegation lines between agents, zoom as navigation, arrow keys on desktop. The classic grid is one toggle away and the automatic fallback without WebGL.
- Team dashboards. Pin a mini-app for the whole team or just for you from any chat, hide a shared app for yourself, ship dashboards inside community agent templates, and a lone dashboard becomes the agent's chrome-free home page. Dashboards use the full width on desktop, and the mini-app kit gained three.js with glow effects.
- Chat. Drag or paste files into the composer with live upload progress; uploads of any size pass strict gateways in ~32 MB parts and resume after a blip; runs of tool calls fold into one summary line (Compact/Detailed setting); file paths in replies open a live preview; typing into a busy Claude chat stops the turn gracefully and answers at once; question cards show each option's full content; notifications open with the sending agent's identity and are capped to headline length.
- Engines and models. Claude Fable 5.1 is the default Claude model, with Claude Code 2.1.258 and Codex 0.149.1 pinned. Pin a single scheduled task to its own model or engine, switch a finished task chat to another engine, and get 72h/24h warnings before a connected Claude or ChatGPT login expires, with a one-click Reconnect; a dead login is reported at once instead of retried forever.
- Skills and tools. Skills can bundle scripts (Anthropic's skill-creator
ships built in); install any standard skill folder from a zip; agents can
browse and request skills and discover every MCP installed on your
platform; a built-in
otodock-platformskill teaches every agent how OtoDock itself works; approving an MCP request lets you pick the instance, and a missing instance shows as "Needs instance" instead of a failure. Agents can write their own persona, and managers are told while it is still empty. - Integrations. Personal access tokens can be saved as separate labeled accounts (a repo-scoped token for one agent next to your own), API-key integrations connect with just the key, and a second Google-powered MCP extends the existing grant instead of replacing it. Excel writes gained real dates, format presets, named-range dropdowns and validation removal.
scripts/sandbox-doctor.shexplains why agent sandboxes cannot start on a host (including the Ubuntu 24.04 user-namespace hardening), which the platform now also detects and names at startup.
Changed
- Delegated workers report back; callers no longer reply by delegation just to acknowledge a result.
- The Groq text classifier is the default end-of-turn judge for every phone-mode language, with the on-box Smart Turn model as the fallback; existing installs keep their stored choice.
- Voice replies keep talking through tool work and start faster; the first-generation buffer is tunable per ElevenLabs provider row.
- Model pricing is refreshed to both vendors' current lists (GPT-5.6 tiers, Claude Sonnet 5 at $2/$10) for cost attribution.
- Satellite auto-update pauses after two rollbacks of the same version; "Update now" retries it.
- Meetings convened by sessions without a user are limited to the agent's delegation targets, and the roster is labeled "Delegation Targets".
- Document text extraction truncates at 512 KB with ranged-read guidance, and every document read, render and write runs in a memory-bounded worker so one pathological file cannot take the host down.
- The "Active now" panels collapse and scroll instead of pushing the page.
Fixed
- Voice conversations: barge-in truly interrupts, including silent tool phases; the transcript keeps the whole utterance; replies no longer stutter, fall silent late in a session or lose your first words; the mic survives long pauses and stays muted when you mute it; a conversation survives its engine process dying; the halo and status follow reality.
- Phone calls no longer go deaf mid-call, crackle, or get cut by false
barge-ins from background noise (
phone_bargein_timer_snow means the minimum speech duration of an interruption). - Dictation in non-English languages no longer drops words already shown.
- Shared libraries: a file deleted in a writable copy no longer comes back, and restoring from the Recover bin into a read-only library is refused.
- Uploads: no endless "Processing…" when no machine needs the file, chunked uploads finish on Docker installs, big batches queue with progress, oversized files say so, huge photos are downscaled instead of dropping the connection, and the photo viewer no longer closes by itself.
- Delegation: worker results are no longer cut off or shown twice, the delegating agent reliably wakes to read them, and usage is attributed to the model that actually ran.
- Sessions: scheduled tasks no longer die at birth or get closed while waiting on a question or a stalled tool; interactive/headless races and duplicate spawns are gone; remote machines shed idle sessions instead of refusing at capacity, forget no ghost sessions after a restart, and run exactly the pinned CLI.
- Task runs no longer report engine failures as successful empty runs, and a task pinned to another engine no longer runs with the wrong model.
- Logins: a transient provider outage no longer expires a healthy subscription, and the "expires soon" warning fires at most twice.
- Permissions: agent sessions see only what their user could (tasks, triggers, notifications, webhook subscriptions), and viewers can no longer replace a team dashboard.
- Integrations: OAuth MCPs that refresh tokens in place start on shared-only agents, Codex keeps MCP URLs with query strings intact, the sign-in page no longer reload-loops with the wake word enabled, and bare-metal installs apply MCP compose changes on restart. Community catalog: github-mcp 1.1.0 fixes schema failures on resumed sessions with Claude Code 2.1.243+.
- Also fixed: Excel named-range corruption and stacked validations, file paths pasted as links reloading the page, stale "Active now" titles, agent settings overflowing on phones, and the 3D map's texture loading, framing on large screens and pinch-zoom.
Security
- A pre-release audit of the newly public surface (remote machines, phone,
five tool sidecars), every finding fixed: sidecars validate the ids they
put into URLs and refuse private-address URLs; skill and MCP installers
reject crafted names, references and compose files; cross-agent file
transfers re-verify paths at write time; a writable library attachment
requires editor access on the source; the Asterisk dialplan filters
caller-supplied variables and Twilio webhooks validate signatures before
spending rate-limit budget; satellites refuse plaintext
ws://to public hosts unless opted in, verify the process behind their Windows control pipe, strip escape bytes from injected text and store their config owner-only. - Remote-machine API responses no longer include credential hashes.
- Dependency updates for the 2026-08/09 advisories:
cryptography50.0.0,aiohttp3.14.3,h24.4.1,click8.3.3,pillow12.3.0,react-router7.18.2,brace-expansion5.0.9.
Upgrade notes (1.4 → 1.5)
- Phone service on an existing install (new installs get it): in the
install directory run
curl -fsSLO https://raw.githubusercontent.com/OtoDock/oto-dock/v1.5.0/docker-compose.phone.yml, appendCOMPOSE_FILE=docker-compose.yml:docker-compose.phone.ymlto.env, thendocker compose up -d. It idles until telephony is set up. - The departments schema migration and the Fable 5 → 5.1 remap of pinned agents, chats and tasks run automatically on first boot. Fable 5.1 needs Claude Code 2.1.251 or newer; paired machines update their CLIs and their satellite on reconnect.
- Asterisk/FreePBX installs: re-paste the generated dialplan snippet.
- The hands-free half-duplex voice mode is gone; installs without the phone service keep dictation only.
- MCP containers on bare-metal installs are recreated once on the first restart if their compose configuration changed.
- Rolling the Codex pin back requires wiping
CODEX_HOMEruntime databases (its migrations are forward-only).
1.4.0
Released on 2026-07-26. The release on GitHub.
Added
- Agents can create agents. Ask an agent to build you a new one — it interviews you, writes the new agent's persona, tools, schedules and onboarding into a template, and installs it with you as its manager. For platform creators and admins; tools the new agent needs that aren't installed yet go through the usual approval queue. On upgraded installs, agents created before 1.4.0 don't get the new tool automatically — enable it per agent from Agent Settings (new agents have it).
- 1GB file uploads and sync. One per-file cap (
OTODOCK_MAX_FILE_MB, default 1024, was 100MB) governs workspace/chat uploads and file sync to remote machines, streamed with bounded memory and written atomically. A new workspace popup shows live upload and per-machine sync progress. Satellites below 0.5.103 keep the old cap until they auto-update. - Claude Opus 5 replaces Opus 4.8 on the Claude Code engine (same price); chats and tasks still pinned to Opus 4.8 are remapped once at startup.
- Continue a finished chat on a different AI engine. Once a chat's session has ended, the model picker offers the agent's other engines — pick one and the chat restarts from its saved history.
- The model picker offers only AI engines you can actually use, and enabling an engine on an agent now requires a platform subscription for it.
- Live "Commands" badges for Codex background terminal commands, matching Claude chats — on remote machines too.
- Rename chats and task runs from the sidebar (task runs can also be deleted, role-gated); clipped titles show in full on hover/long-press.
- Community agent templates can bundle Agent Skills, and any agent can
ship a per-user onboarding guide (
config/user-setup.md) that loads into each new user's chats until that user completes it. - Workspace quality-of-life: type-ahead selection, per-section recursive search, right-click Refresh, and a clear warning when deleting files too large for the Recover bin.
Changed
- The agent persona file is now
config/agent.md(wasprompt.md) — existing agents migrate automatically at startup. Rolling back to 1.3.x afterwards needscp config/agent.md config/prompt.mdper agent. - Interactive terminals follow the CLI's own permission modes (Shift+Tab) instead of the platform's generic approval prompt; the platform still enforces file-access rules and prompts for the highest-risk actions. Headless and "Don't ask" chats are unchanged — see docs.otodock.io/security/permissions.
- Core tools are granted when an agent is created or installed, and no
longer silently re-added on every start — removing one now sticks.
Templates can opt out entirely with
"core_mcps": "none". - Agent-settings and MCP-marketplace tools no longer load in scheduled tasks or meetings — only where a person is present and asking.
- AI engine CLIs upgraded (Claude Code 2.1.220, Codex 0.145.0). Codex costs now include prompt-cache write tokens (previously undercounted), and GPT-5.6 context windows are corrected to 272k.
- Codex sessions on machines paired with full filesystem access are no longer confined to the workspace by Codex's own sandbox; the platform's permission prompts and path policy still apply.
- Task-history deletion follows the task role matrix, and on shared-only agents viewers can no longer delete shared chats.
- AI chat titles appear much sooner on long, tool-heavy first turns.
Fixed
- Chat stability: no more self-scrolling when document previews load, long chats scroll inside the chat again, and the "Connect an AI engine" banner no longer covers the top bar.
- Mini-apps and artifacts: action buttons can call every platform tool, external links open properly (after a one-time consent), and apps written as complete HTML documents regain theming, actions and live data.
- Tools approved from Agent Settings now start automatically, and saving changed tool credentials restarts the tool's container by itself.
- Document tools: page screenshots return in seconds instead of timing out, scanned PDFs are read with vision (OCR gains a language option and much smaller output), and PDF page export no longer fails on files it just wrote.
- Codex: sandboxed commands work in personal chats, local model endpoints (Ollama/LM Studio) connect again, and permission hooks run on Windows.
- Also fixed: relative paths in terminal commands, opening a freshly granted agent without a page reload, agent managers without a platform role saving configuration, agent renames showing in chat, YouTube links embedding again, and the memory tool recovering from malformed calls.
Security
- On a shared agent, a colleague's live terminal now opens read-only with a Take over button instead of accepting your input into a session running under their account and permissions; terminal messages are attributed to whoever typed them.
- Updated
pyasn1to 0.6.4 (two denial-of-service advisories).
1.3.2
Released on 2026-07-23. The release on GitHub.
Added
- The installer now offers a one-time host tuning (
vm.swappiness=10via/etc/sysctl.d/, sudo-gated, safe to decline) that prevents dashboard stalls after memory-heavy tool runs. Existing installs can apply it once withsudo bash scripts/setup-host-tuning.sh.
Fixed
- Viewing a generating chat from more than one device or tab no longer makes the view reload every few seconds — all viewers now stream the turn live, simultaneously. This also removes the response lag those reloads caused.
- The chat no longer flashes a reload right after a response completes, and sending a message to a live interactive terminal from the chat input no longer reloads the terminal view.
- Hardened the Windows shell-wrapper analysis in the command permission gate against pathological regex backtracking on adversarial command strings.
- SSO login no longer fails with "OIDC not configured" until a proxy restart when the identity provider was unreachable at proxy startup (e.g. proxy and a co-hosted IdP cold-starting together after a power cut). OIDC endpoint discovery now retries automatically on the next login attempt, at most once per 30 seconds; explicitly configured endpoint URLs are never overwritten.
- Admin accounts now see only their own personal folder in the agent workspace file browser, like managers — this also fixes admins sometimes being shown an empty "My Workspace".
- Workspace sync to a user-paired remote machine now runs with the owner's
per-agent role instead of full admin authority, and agent
config/files are never deleted through sync absence-inference anymore (deliberate deletes still propagate). Previously a machine that had lost its working copy could delete the agent's prompt on the platform at sync time. - Deleted personal files in the recover bin are now restorable only by their owner; shared workspace/knowledge/config entries keep their editor/manager tiers.
1.3.1
Released on 2026-07-21. The release on GitHub.
Fixed
- The Remote Machines settings tab no longer appears after a fresh login on
builds that ship without the remote-machines feature. Every login response
(password, 2FA, passkey, OAuth, account creation) now carries the same
feature flags as
/auth/me, so the dashboard sees the correct feature state immediately instead of only after a page reload. - Idle session cleanup works from the moment the server boots: sessions with no recorded activity were mistakenly held open for the first 15 minutes after a host restart.
1.3.0
Released on 2026-07-21. The release on GitHub.
Upgrade note (Docker installs from before 1.2.0): the compose file now reads operator settings from a
.envfile next todocker-compose.yml(older installs usedconfig.env). If your settings still live inconfig.env, rename it —mv config.env .env— or setOTODOCK_ENV_FILE=config.env. Starting without a.envfile no longer crash-loops: the proxy detects Docker's silently-created directory and stops with instructions naming the exact fix.
Added
- The video toolkit. Agents can now produce finished videos end to end. Three MCPs ship with the platform: video-gen (AI footage, transitions and Runway Aleph edits, with your Google / Runway / fal.ai keys), music-gen (ElevenLabs music and sound effects) and tts (voice-overs via the platform's TTS providers). The video-tools editing MCP (timelines, transitions, captions, color grading, FFmpeg rendering) is in the community catalog, installable on any 1.0+ install.
- One-command install.
scripts/install.shbootstraps a fresh Docker install end to end: Docker preflight,.envgeneration, the Ubuntu 24.04+ AppArmor step, and first start. - Chats tell you where they run: a chat stays on the machine it started on, says so when that differs from the agent's current machine, and an owner/admin can move it to the current machine with its conversation reloaded.
- Delegation: a session can adopt an existing project as its orchestrator, and the delegation tool lists each agent's layers and models so requests are validated instead of failing blind.
- The SSH MCP can list its authorized hosts mid-session (
list_ssh_hosts).
Changed
- Interactive terminal sessions are enabled by default, and tool permission prompts are now risk-based: read-only tools never prompt, reversible actions stop prompting in Accept Edits mode, and outward-facing or costly actions always prompt. MCPs without a declared tier keep their old prompting behavior.
- The installer installs into the directory it is run from, refuses your home directory, and performs fresh installs only, as before.
- Per-agent roles decoupled from the platform role: any non-admin user can hold any per-agent role (manager / editor / viewer).
- Shared-only agents bill the person, not the platform: dashboard chats are attributed to the user whose subscription serves the session, and a shared chat changing hands recycles the live session under the new sender so one user's messages never spend another user's account.
- Remote agents allowed to run
sshcan read~/.ssh/configandknown_hosts(private keys stay unreadable, writes stay denied).
Fixed
- Interactive terminals: the first message into a cold terminal is no longer lost, assistant text no longer duplicates, a dying CLI says "process exited" instead of going blank, opening a terminal no longer overwrites your clipboard, and Codex permission questions show their card in headless mode.
- Remote machines: replies no longer land one turn late after background work, the first workspace sync shows live progress instead of silent minutes, zero-byte files and folder deletions sync correctly, files on the machine's own disk preview and save reliably, and MCPs that need a newer Python install it instead of failing forever.
- Delegation: results reliably wake sleeping orchestrators (and are replayed if delivery fails), stopping a worker mid-turn no longer kills its session, and the "Delegated" badge always completes.
- Documents: previews in older chats no longer show "session expired", and the setup wizard pins document preview to the address you browse from, so fresh installs get a working preview on any origin.
- Tasks: "Run now" shows an honest queued state and attaches the open chat
when streaming starts;
run_taskwithwait: trueworks again. - Connected Claude accounts keep their plan tier (Max/Pro) across token refreshes, so sessions no longer refuse plan-included models.
- Voice input no longer appends an invented filler word after you stop speaking.
- Plus a round of smaller fixes across meetings, chat deletion, skill loading, passkeys, browser-tool contention, and platform shutdown.
Security
- Interactive terminal sessions enforce the chat's permission mode: in Default mode, file edits, gated shell commands and tool calls prompt in the terminal before running; mode changes made inside the terminal are respected.
- Display and media hooks are confined to the agent workspace, closing a path that could read arbitrary proxy-host files into the chat.
- OAuth connect pages escape reflected input and confine post-login redirects; API errors return clean messages and keep exception detail server-side.
- Template and text regexes are linear-time (ReDoS), the per-session Codex config is written owner-only, and prompt-file loading is confined to the agent tree.
- Dependency refresh across the proxy and bundled MCP servers — every open Dependabot advisory with an available fix is cleared.
1.2.0
Released on 2026-07-18. The release on GitHub.
Added
- Agent Skills. Skills now follow the industry-standard Agent Skills format
(SKILL.md, agentskills.io) and load two ways:
always(inlined into the system prompt) oron_demand(loaded only when a task matches — most skills moved there, slimming every system prompt). Standalone skill packages install from the new community-skills catalog — seeded with curated Anthropic and OpenAI skills — with the same admin approval flow as MCPs, managed from a new Skills tab in Agent Settings and a Skills page in admin settings. Docker installs gain a named volume (otodock-skills), created automatically bydocker compose up. - Real equations in documents. Agents write LaTeX math into Word (native, Word-editable), PDF (vector-rendered), PowerPoint and Excel (high-resolution images that persist across later edits), and read it back out as LaTeX. sympy now ships in the baseline agent toolchain — existing hosts pick it up on the next installer run or image rebuild.
- Previous-version document previews. When an agent delivers a new version of a document it already previewed, the old preview stays on screen as a view-only "previous version" you can scroll back to and compare; only the newest preview is editable, and a file keeps at most two full previews.
Changed
- Spreadsheet reads return a coordinate-labeled grid and every write reports back the cells it touched — fixing a class of answers-one-column-off errors. Fresh formulas show as formulas, and malformed operations are reported instead of silently skipped.
- Skill context exclusions now actually apply: scheduled tasks and phone calls no longer load skills whose manifest excludes those contexts.
- Repeated reads of the same remote-machine file are now instant: the platform first asks the machine whether the file changed (satellite 0.5.95 — connected machines self-update) and serves its cached copy when it hasn't. Any doubt still transfers the full file.
Fixed
- A connected account revoked at the vendor no longer retries its token refresh every minute forever — refreshes back off, and a definitively revoked account waits until you reconnect it.
- Document preview: now works on bare-metal installs served over plain HTTP or a non-default port ("refused to connect"), survives Collabora stalling on heavy spreadsheets, keeps an idle document loaded for 2 hours (was 1), and a view-only viewer can no longer place an editing lock that blocked other users' saves.
- The shared local browser no longer opens by itself: the window now appears only when an agent actually runs a browser action, and stays closed once you close it.
- Interactive terminal: content no longer goes missing after sitting idle, no keypress is needed to revive scrolling, switching chats no longer leaks a GPU context, and a prompt sent while the terminal was still starting is delivered once it's ready instead of silently dropped.
- Task runs opened from the sidebar now show the run's real model and permission posture instead of the viewer's own selections — and a task's "Don't Ask" mode no longer leaks into the next new chat you open.
- A scheduled task whose agent left a background command running forever no longer loops endlessly re-reviewing it — the task waits a bounded time and completes, noting anything still running.
- Remote machines: a newly created, deleted, or catalog-installed agent shows up in Remote Machines without a page reload; uploads to remote agents no longer stall for the length of the transfer (the copy continues in the background and a prompt referencing the file waits for it to land); and document reads over slow connections get the platform's full 150 s transfer window instead of failing at 5 s.
- Community MCP containers self-heal after the install's identity changes (stale container names are re-stamped and the start retried); containers from a different OtoDock install are named in a boot warning, never touched.
- Silent failures now fail loud at boot: one ERROR names the cause and fix
when stored credentials can no longer be decrypted (changed
JWT_SECRET/config.env), another lists the exact columns when the database schema is missing ones the code expects. - Docker installs on Ubuntu 24.04+ no longer fail at boot with a namespace
error. OtoDock ships a scoped AppArmor profile (
otodock_userns) that grants the needed capability to the OtoDock container only — instead of the widely-circulated workaround that disables that kernel hardening system-wide.scripts/compose.shinstalls it automatically (one-time sudo); pull-only installs runscripts/setup-apparmor-userns.shonce and setOTODOCK_APPARMOR_PROFILE=otodock_usernsin.env.
1.1.1
Released on 2026-07-16. The release on GitHub.
Changed
- Chat dictation can now run up to 3 minutes per take (was 60 seconds, which
cut long dictations off mid-sentence). Admins can tune this with the
audio_chat_stt_max_secondssetting.
Fixed
- Live voice mode no longer sends truncated turns. The auto-send fired the moment the silence timer expired, racing the speech provider's own end-of-utterance commit — long turns went out with the tail missing and anything said after was lost. The mic now stops first, the provider flushes everything it heard, and the full transcript is what gets sent. Stopping dictation likewise waits briefly for that flush, so the last sentence is never dropped.
- Dictation no longer doubles your sentences. Two ways a hidden second recording session could be left running — a slow microphone connect timing out, and stopping the mic while it was still connecting — meant a retry had multiple sessions transcribing the same microphone, so every sentence landed in the input twice (or more, stacking with each retry).
- The sidebar's "Active now" strip shows the real titles of live chats (it showed "New chat" for chats that started after the page loaded — most visible in the task-history view).
1.1.0
Released on 2026-07-15. The release on GitHub.
Added
- Images opened full-screen from a gallery can now be zoomed: pinch on touch, mouse wheel or double-click on desktop, with panning while zoomed. Long captions are shown in full.
Fixed
- Video fullscreen on mobile no longer pins the screen to landscape: wide videos still rotate as a starting cue, but turning the phone upright rotates the video back to portrait fullscreen. Vertical videos are no longer rotated to landscape when fullscreen starts before the video dimensions are known.
- ElevenLabs speech-to-text no longer invents trailing text after you stop the chat mic. Stopping after a pause committed a silence-only buffer, which the model would "transcribe" into words never spoken (most noticeable in non-English dictation).
- A speech-to-text provider that fails (bad API key, quota, connection) now shows an error on the chat mic instead of a mic that hears nothing.
- Audio settings: the provider API-key row (Save/Remove buttons) no longer overflows the screen on mobile.
1.0.2
Released on 2026-07-14. The release on GitHub.
Added
- This changelog. Every release from here on ships its notes with it, so you can see what changed before you pull.
Fixed
- The proxy reported version
1.0.0in its API documentation regardless of which release was actually running. It now reports the running version.
1.0.1
Released on 2026-07-13. The release on GitHub.
Fixed
- Signing in could land you on a dead page. The redirect from
/resolved against the wrong list of agents, so anyone who had agents but had not picked a favourite hit a broken "Back to Chat". It now falls back to the first agent you can actually see, for every role.
1.0.0
Released on 2026-07-13. The release on GitHub.
Initial public release.
OtoDock is a self-hosted platform for running a team of AI agents on infrastructure you control. It runs the real Claude Code and Codex as its engine, so your agents inherit everything those CLIs can do, and wraps them in a live dashboard, a security model built for shared servers, and the plumbing that turns a coding tool into a team of coworkers.
Added
- Agents and chat. Every step streams live — reasoning, tool calls, file edits as diffs, plans ticking off. Approve sensitive actions inline, or let trusted agents run unattended.
- Multi-agent meetings. Put specialists in one room for a moderated discussion where agents address each other, answer in parallel, and converge.
- Delegation. Agents hand work to parallel agent sessions you can watch, steer, and continue.
- Bring your own engine. Connect the Claude or ChatGPT plan you already pay for, an API key, or a local model.
- Sandboxed by default. Each agent runs locked down and isolated from your network; you grant one folder or one service at a time.
- Schedules and triggers. Recurring and one-off background tasks, plus webhooks that let outside systems start work.
- Persistent memory that survives across sessions.
- Documents and images. Read and author Word, Excel, PowerPoint and PDF, with a live in-chat preview; edit and generate images.
- Voice. Speak to your agents and have them answer out loud.
- Interactive artifacts and pinned mini-apps — dashboards and small tools your agents build and you keep.
- Community catalogs for installable MCP capabilities and ready-made agent templates.
- Self-hosted install via Docker Compose, with your chats, files, memory and credentials staying on hardware you run.