Repository navigation
v1.1.1
Before upgrading. Read these first: several change what a Helm values file, a Compose override
or a script has to say.
- The Helm charts refuse shared and external computers until you say you mean them. Shared
computers (computers.mode: shared, still the default) and external ones
(computers.mode: external, the one computer atcomputers.url) run every person's Bots on one
computer, where a command one Bot runs can read another person's files, so they suit a proof of
concept only. Sandbox mode (computers.mode: sandbox) gives each person a computer of their own
and is the way to run OpenBot for people; it needs the agent-sandbox controller in the cluster.
A release on shared or external computers, including one that never setcomputers.mode, has to
setcomputers.sharedModeAcknowledged: truebefore upgrading
(openbot.computers.sharedModeAcknowledgedin theopenbot-bundlechart), orhelm installand
helm upgradestop with a message naming both choices and change nothing. That includes an
upgrade with--reuse-values, and the single-user mode of either chart. Acknowledged, the
install notes say it again, and admins see it in Admin (below). - The
openbot-bundlechart keeps files in its own object store. A bundle upgraded with your
own values file moves every file people and Bots attached out of PostgreSQL and into the bucket
openbot-filesin the bundled Intelligence's local object storage, which a new Job makes on
every install and upgrade. That storage is one pod, with no replication or backup of its own. To
keep files in PostgreSQL, setopenbot.attachments.store: postgresand
openbot.objectStorage.enabled: false. An upgrade with--reuse-valueskeeps the values it had,
so files stay in PostgreSQL. Don't turn offintelligence.localObjectStoragewhile
openbot.objectStoragepoints at it: the bundle refuses that. See Files (below). - Migrations run at start: five, all additive. The database moves from
0132_bot_task_runsto
the migrationrelease-policy.jsonnames for this release.0133_files_in_an_object_storeadds
thefile_deletionstable and lets an attachment's bytes live outside PostgreSQL,
0134_attachments_storage_key_indexindexes where they live,0135_files_gate_openedrecords
when files may move,0136_api_deprecated_usecounts use of deprecated API, and
0137_ai_connectionskeeps the AI connection an admin chose (below). A 1.1.0
server works against the new schema, so 1.1.0 upgrades to 1.1.1 with its servers rolling one by
one, with no downtime.0133needs OpenBot's login to hold theopenbot_systemrole's
privileges, as 1.1.0 set up, and stops with theGRANTto run when it doesn't. It waits at most
five seconds for a lock onattachments, so a long transaction can fail the first start, which
then tries again. Upgrading from 1.0.x or 0.3.x runs 1.1.0's migrations as well: read 1.1.0's
Before upgrading first. - Going back to 1.1.0 after files moved to an object store needs them moved back first: set
attachments.store: postgreswith the store still configured, and wait for the move to finish.
A 1.1.0 server cannot read a file that is in the store. - New Bots run on the LangGraph (TypeScript) engine where a deployment runs it. OpenBot
Desktop runs it, and an installation's Bot that ran on the Python LangGraph engine moves to the
TypeScript one at its next Install and Start, keeping its conversations. Docker Compose now
definesagent-langgraph-ts(port 4215, on this machine only), which a plaindocker compose up
builds from the checkout unlessLANGGRAPH_TS_IMAGEandIMAGE_PULL_POLICY=missingname the
published image. A server withoutLANGGRAPH_TS_AG_UI_URL, and both Helm charts, keep new Bots
on the server's built-in engine.scripts/start.shstarts it on 4215 and sets
LANGGRAPH_TS_AG_UI_URLto it, except in the personal edition (LANGGRAPH_TS_PORTmoves it).
OPENBOT_DEFAULT_BOT_ENGINE=built-inkeeps every new Bot there. - Compose: Claude and ChatGPT plans run in services of their own.
agent-harnessis no longer
handedCLAUDE_CODE_OAUTH_TOKEN,CHATGPT_AUTH_FILEor the./.langchainfolder. They go to
two new services,agent-claude-planandagent-chatgpt-plan, behind theclaude-planand
chatgpt-planprofiles, which OpenBot Desktop starts when a plan needs them. A Compose override
of your own that ran a plan throughagent-harnessturns on the profile and sets
CLAUDE_PLAN_HARNESS_URLorCHATGPT_PLAN_HARNESS_URL.agent-langgraph-aguireads the ChatGPT
plan's./.langchainfolder too, and.env.examplepointsCHATGPT_PLAN_HARNESS_URLat it. - The model picker offers only models proven on the Bot's engine (below). A conversation or a
Bot set to a model that isn't proven runs on the default, and says so. - A model gateway has to serve
/responsesfor gpt-5.4 and gpt-5.5. Bots behind an
OpenAI-compatible address now call those models, and their mini, nano and dated versions, on the
Responses API, because they refuse tools with a reasoning effort on Chat Completions.
COPILOTKIT_OPENAI_API=chatkeeps them on Chat Completions, for the built-in Bots and the AG2
and Langroid harnesses. - A server that refuses to start exits with code 78, where it exited with 1. The all-in-one
image now ends its container instead of trying again forever, and on Kubernetes the pod's
CrashLoopBackOff names the reason inkubectl describe. A crash still exits with 1. Monitoring
that watches the exit code may need to know the new one. - Settings that now stop a server from starting when they are wrong:
OPENBOT_MINIMUM_CLIENTS,
OPENBOT_DEFAULT_BOT_ENGINE,SHUTDOWN_KEEP_SERVING_SECONDS,SHUTDOWN_DRAIN_SECONDS,
RELEASE_STATUS_PORT, and the file store's (below). AnS3_BUCKET,GCS_BUCKETor
AZURE_STORAGE_CONTAINERalready in the server's environment, for example one it shares with
Intelligence, configures a file store even withFILES_STOREunset, and the server then checks
that store at start and refuses to start when it can't use it. - Kubernetes: a server pod takes up to 45 seconds to stop, so it can finish what it is doing
(below).server.shutdownsets it. - The
openbotchart needs this release's image. Its wait for the database now runs from the
server's own image (below), so the new chart with an olderimage.tagdoesn't start. - Deprecated: nothing.
Files
- Files can live in an object store. Amazon S3 or anything that speaks its API, Google Cloud
Storage or Azure Blob Storage, where they were always in PostgreSQL. With a store, a file can be
up to 100 MB. Without one nothing changes: files stay in PostgreSQL, with 8 MB for pictures and
PDFs and 1 MB for other files, and OpenBot Desktop stays on PostgreSQL. Theopenbotchart gains
attachments.store(postgresunless set) andobjectStorage:enabled,provider(s3,
gcsorazure),bucket,prefix(openbot/),region,endpoint,forcePathStyle,
azure.accountUrl,deleteAfterDays(7),assumeAllServersCurrent,existingSecretwith its
secretKeys, andinClusterfor a store in the cluster. A server pod can sign in to the store
with workload identity throughserviceAccount.annotations. On Docker Compose the same settings
areFILES_STORE,FILES_PREFIX,FILES_DELETE_AFTER_DAYS,FILES_ASSUME_ALL_SERVERS_CURRENT,
S3_*,GCS_BUCKETandAZURE_STORAGE_*in.env. The chart refuses a store asked for and
not set up, a Secret without its keys, and a store in the cluster behind a network policy with no
way to reach it, each by name.- The server writes, reads and deletes a test file at start, and refuses to start when the store
doesn't answer. It never falls back to PostgreSQL. - Files already in PostgreSQL move to the store in the background, and back again when
attachments.storeis set topostgreswith the store still configured. A server refuses to
start without the store's settings while any file is still in it. - During a rolling upgrade new files stay in PostgreSQL until no 1.1.0 server is connected. A
connection pooler, such as PgBouncer or RDS Proxy, hides which servers are connected: set
assumeAllServersCurrent: trueonce every server runs 1.1.1. - A deleted file leaves the store after
deleteAfterDays, and a weekly sweep removes anything
left behind. Files are served from OpenBot's own addresses, as before.
- The server writes, reads and deletes a test file at start, and refuses to start when the store
- Word, Excel and PowerPoint files. People can attach
.docx,.xlsxand.pptxfiles, and
Bots can send them, recognised by what is inside rather than by the name. The Bot reads their
text: a workbook's sheets as tables, a document with its tables, a presentation's slides in
order, up to 120,000 characters. They count as documents, so the document limit applies. Old
.doc,.xlsand.pptfiles, password-protected ones, macro-enabled ones (.docm,.xlsm,
.pptm), templates, and a file whose contents don't match its name are refused with a sentence
saying why. - Bots send files to people. A Bot can hand you a file from its computer. It shows as a card in
the conversation, a task or a group, with Download, and is still there after a reload. A link to
the file in the Bot's words downloads it, where it opened the home page. - The limits come from the server. The message box refuses a file over this OpenBot's limit
before uploading it, and a picture over 8 MB shows as a file. A picture over the model's own limit
reaches the Bot as a note instead of failing the turn.
Apps and the OpenBot they talk to
- Every app says which version it is, and OpenBot can ask for a newer one. The web app, OpenBot
Desktop, the phone app and the Recorder sendOpenBot-Client: <app>/<version>on every request,
and the server'sdeploymentanswer names the API versions it serves and the oldest release of
each app it serves (api,minimumClients,launcherApi,extensionApi). An app older than
that reads "Update OpenBot" in place of the app, and a server too old for the app says "Ask
whoever runs your OpenBot to update it". A server refuses an older app's requests with 400
CLIENT_UPDATE_REQUIRED. Nothing is refused unless an admin setsOPENBOT_MINIMUM_CLIENTS
(for exampledesktop=1.2.0,mobile=1.2.0), because a release serves every app released in the
12 months before it. A request with noOpenBot-Clientheader, such as a script's, is never
refused. - OpenBot Desktop 1.0.3 or older gets a page instead of a server that stops. An older Desktop
that starts this release now shows "Update OpenBot" with Check for Updates…, where the server
used to exit for want of the install token those apps never wrote. - OpenBot Desktop checks it can start a release before it starts it. A release that needs a
newer app is not started; the window offers Update and restart instead. - A signed security floor. The apps, and OpenBot Cloud's servers, can be told the lowest safe
release of OpenBot Desktop, its back end, the phone app and the Recorder; an app below it shows
one notice, "This version of OpenBot has a security problem. Update to keep using it." It never
applies to a company's own OpenBot, and no floor is set
(docs/releasing.md).
API
- The API refuses NUL in text. Every text field the spec holds to a pattern refuses a NUL
character anywhere in it, with 400VALIDATION_ERRORand "That text has a character OpenBot
can't keep. Type it again." Before, one NUL could pass where the pattern asked for a character. - The API says when something you call is deprecated. An operation, a query or header
parameter, or a request-body field the spec marks deprecated answers with the standard
DeprecationandSunsetheaders and aLinkto this changelog, whose Deprecated list under
"Before upgrading" says what replaces it. It keeps working until at least its sunset, 12 months
or more after it was deprecated. Nothing is deprecated yet, so no answer carries them. See
api/README.md. - Statuses, roles and scope kinds in API answers can gain values. The spec declares them open
(x-extensible-enum): a task's or a run's status, a person's role, a network policy's scope,
what decided a person's permission, and the others listed inapi/accepted-breaking-changes.md.
A script that reads one should treat a value it doesn't know as "other" rather than fail, as the
web and phone apps now do. The values sent today are unchanged. - Edit a routine in place.
PATCHon a routine changes its name, what to do, its schedule and
its time zone, keeping its conversation and runs, and a routine's page has an Edit routine row.
Before, the only way to change one was to ask its Bot. A routine in an answer now always carries
itscron. - New optional fields:
computerModein the organization settings'context,attachmentLimits
andsecurityMinimumClientsondeployment,filesanddeclinedon a message, andunverifiedon a conversation's model
choice. A Bot'sframeworkcan belanggraph-ts.
Bots and models
- Change the AI connection with no restart. An admin of a personal OpenBot changes the AI
connection from Settings > General > AI model: a key, an address, or, in OpenBot Desktop, a plan
or provider sign-in. OpenBot checks it with the provider before saving it, and the next run uses
it. A team's OpenBot keeps it as its deployment sets it: Settings doesn't offer it, andPUT
answers 409AI_CONNECTION_FIXED.
GETandPUT /api/v2/organizations/{organizationId}/ai-connectiondo the same, and with
MODEL_GATEWAY_TOKENset the Bots' engines reach models through OpenBot instead of holding a
provider key of their own. - The model picker offers a model only when OpenBot has proven it on the Bot's engine. A daily
check sends every listed model a request through each engine, and a model that doesn't answer
there is hidden, on OpenAI, Anthropic and Google. A conversation already set to one runs on the
default, and the composer says "X doesn't work with this Bot yet, so this conversation uses Y".
A provider never checked, such as xAI, and a gateway, list everything as before. - New Bots on LangGraph (TypeScript). See Before upgrading. A Bot's Setup > Connection says
what it runs on. - Claude plan errors say what happened. A usage limit, running out of credit, a dropped
connection, no answer, a client too old for the model and a refusal each get their own sentence,
and a task tries again after a dropped connection but not after a limit. A call to a tool the
step didn't offer goes back to the model as a tool error instead of ending the turn, and a reply
that came back cut off says "Ask it to carry on". The Claude harness bundles a newer Claude Code,
so plan Bots run Claude's newest models. - Gemini sign-in says what Google said, with when to try again. A free-tier daily limit reads as
a rate limit, never as out of credit, and the model list loads when OpenBot starts. - A failed turn that had already done something keeps its message sent. On every engine, when
the Bot had said something or called a tool, the message stays sent and the failure shows under
the turn, instead of coming back in the message box to run again from the start. A turn that
never got going still gives the words back. - OpenBot reaches a Bot you run yourself, a plan's engine and MCP servers on a connection of its
own, so a second message a few seconds after the first no longer fails with "Couldn't reach"
the Bot. A built-in
Bot's tool calls are no longer lost under load. - The AG2, Langroid and LangGraph harnesses call gpt-5.4 and newer, and the -pro and -codex
models, on the Responses API;COPILOTKIT_OPENAI_API=responsesorchatchooses one for every
model in AG2 and Langroid.
The Agno harness no longer sends Claude atop_p, which Claude 4.5 and 4.6 refused. The
LangGraph harness keeps the kind of error Gemini and xAI name, and pins its model clients.
Conversations
- Group conversations name who sent each line, by name or email, and "Someone who left" for
someone no longer there. - A Bot's reply waiting to be shared says whose turn it is. In a group, only the Bot's owner sees
Open on it; everyone else reads "Waiting for to allow sharing this reply." A reply the
owner declined to share reads " didn't share this reply. Nothing was sent." as a notice,
not an error. - A Bot with nothing to say in a group stays quiet. A message that names no Bot goes to every
Bot in the group, and one that finds it isn't theirs to answer now leaves no line, where everyone
saw "The turn finished without saying anything." in red under it. A retry doesn't run it again. A
Bot that was addressed (a chip, an@and its name, a handoff, an answer to its question) and says nothing
still fails, as in a one-to-one conversation. - Dictation in OpenBot Desktop on a Mac falls back to the next language and region the Mac has
when the page's isn't available, and says which to add. A browser says when it can't take
dictation in a language.
Admin
- Admins see when everyone's Bots share one computer. On a team's OpenBot whose Bots all run on
one computer, Admin's overview and its Computers page say so, say it suits a proof of concept
only, and say how to give each person a computer of their own. An OpenBot one person uses, OpenBot
Desktop included, says nothing, because that computer is the person's own. Nobody but an admin
sees anything new. The organization settings' read-onlycontextgainscomputerMode
(per-person,sharedornone).
Personal edition
- OpenBot for one person asks once to be starred on GitHub. In the personal edition only, once
you have used OpenBot on three separate days, a card above the message box asks you to star
OpenBot on GitHub, right after something goes well: you copy a Bot's answer, one of your
routines finishes a run, or a long task finishes. Never right after an error, never on a team's
OpenBot, and never beside the notifications offer. Either answer is saved to your account, so it
is never asked again on any device. Settings > General keeps the link in a new OpenBot section.
The preferences gaindismissedNotices(the notices you answered) andnoticesDue(the ones
OpenBot would show you now), andPOST /api/v2/organizations/{organizationId}/nudge-events
counts how many people were shown a nudge, followed it or dismissed it, with product telemetry
on; see TELEMETRY.md.
OpenBot Desktop
- Connected to another OpenBot on a Mac works again. In 1.1.0 the window showed only "This
relay is for the OpenBot window", because the Mac's web view held back the sign-in cookie after a
redirect. - A server that refuses to start says why at once, in its own words, where the window waited
three minutes and said it could not finish starting. Retry after fixing it starts clean, without
the old refusal. - Retry waits for a back end whose services are still coming up. When the server had stopped
and some of the back end's services were still starting, Retry gave up on its first look and
said OpenBot was running but not working properly. It now waits for them as it waits for a back
end that is starting, and says that only if they are still not ready at the time limit. - Stop and Quit let the server finish first. The server stops before the Bots and computers it
is using, and a Bot's answer in progress ends as stopped, not as "Couldn't reach". Stop and Quit
no longer cut each other's stop of the back end short, so Quit during a Stop on Windows no longer
leaves it running. Stopping can take a few seconds longer. - An update checks this computer has room first, before it stops or downloads anything, and
says "OpenBot needs more free space to update" with the figures and Retry. A full disk reads as
one plain line. - Status window: Restart applies changes made to
.env, it offers Start after Stop, and it
shows Updating with no buttons during an update. - Launching OpenBot while it quits opens it again once it has quit, where it stayed on Shutting
down. - The first Start after Install waits up to two minutes for the container engine, instead of
sending you back to Install. - On Windows, checking your saved setup no longer freezes the window while Windows finishes an
update, and says so after five seconds. - An update stopped after it updated your data says it didn't finish and that the next Start
finishes it. An update to the release OpenBot already runs is not a failure. - The first screen offers OpenBot (Self-Contained), and one OpenBot for Business card, Cloud or
Self-Hosted, that books a call; connecting to an OpenBot at work is "Already have OpenBot at
work? Connect to it." The update screen says the same. - The step before a Composio sign-in shows inside the Desktop window.
- Get Troubleshooting Logs adds the speech log and the model provider's own error words, with
secrets taken out.
Running OpenBot
- A server asked to stop finishes what it is doing. It finishes requests and Bot answers in
flight, closes live connections with 1001, and keeps an answer still running at the limit as
stopped.SHUTDOWN_KEEP_SERVING_SECONDS(0 unless set) andSHUTDOWN_DRAIN_SECONDS(6) set how
long; the chart'sserver.shutdownuses 5 and 25, and a pod'sterminationGracePeriodSeconds
becomes their sum plus 15. A second signal stops it at once. - A browser or database switched off in the OpenBot image stays off
(EMBEDDED_COMPUTER=off,EMBEDDED_POSTGRES=off, as the charts' server pods set unless
server.embeddedComputeris on), where it was
restarted about once a second. A computer that exits no longer leaves its virtual display
running. - The
openbot-bundlechart installs on Amazon RDS and other managed PostgreSQL. There only the
provider's extension role (rds_superuser,cloudsqlsuperuser,alloydbsuperuseror
azure_pg_admin) can create pgvector, so on a database server of your own the bundle's setup Job
gives OpenBot's role that extension role only while its migrations run on an empty database, and
takes it back once they are done. The Job now also runs after each install and upgrade to do that;
withhelm install --waitit is the run straight after the install. A new
externalDatabase.intelligenceParametersgives Intelligence connection parameters of its own,
where its driver readsparametersdifferently from OpenBot's: on RDS, Intelligence needs
sslmode=verify-full&sslrootcert=/etc/cpki/rds-ca/global-bundle.pem, and its images carry RDS's
certificate bundle. Unset, it takesparametersas before. - A database certificate checked against a CA of your own works at start. With
sslmode=verify-full&sslrootcert=<path>, as Cloud SQL and Azure Database need, the server's
migrations now check the same CA its connections do, where they failed with "self signed
certificate in certificate chain" and the server never started. In theopenbot-bundlechart,
setup.extraVolumesandsetup.extraVolumeMountsmount that CA for the setup Job, and on Azure
Database the Job lends Intelligence's roleazure_pg_adminuntil its pgvector migration has run,
then takes it back. Withglobal.storageClassset, the bundle makes the claim for Intelligence's
object store with that class, so a cluster with no default StorageClass, such as a new GKE
Autopilot cluster, installs. - The
openbotchart's server waits for its database with nothing downloaded. The init
container that holds the server until the database answers now runsopenbot wait-for-database
from the image, with the server's own database client and TLS settings, and gets
config.extraVolumeMountsas the server does. In 1.1.0 it fetched a PostgreSQL client from the
npm registry in every pod, so on a cluster with no way out to the internet the server never
started. It needs this release's image: the chart with an olderimage.tagdoesn't start. helm install --waitandhelm upgrade --waitfinish under Helm 4 and Flux. The charts'
Jobs are no longer deleted on a timer after they finish, which Helm 4 read as never ready,
marking the release failed on a slow cluster. Each stays until the next upgrade replaces it.- Signed charts in a registry. Every release publishes the
openbotandopenbot-bundle
charts as signed OCI Helm charts,oci://ghcr.io/copilotkit/charts/openbotand
oci://ghcr.io/copilotkit/charts/openbot-bundle, at the release's version, signed with the key
indocs/cosign.pub. The release'sopenbot-release.jsonnames each digest. Check one with
cosign verifyand install it by digest with no checkout, or let a FluxOCIRepositoryverify
it. - Release and health from inside the cluster.
/releasenow names the database's newest
migration.RELEASE_STATUS_PORTadds a second port that answers only/releaseand/metrics
(requests, errors and latency per route, and whether the database answers), so a collector needs
nopods/execon the server. In theopenbotchart,server.internalStatus(enabled,port
9464,allowFrom) puts it on a ClusterIP Service of its own,<release>-server-internal,
behind a NetworkPolicy that admits only those peers. ARELEASE_STATUS_PORTthat isn't a port
number, or is the API's own port, stops the server from starting.
Smaller changes
- An admin's review list shows a person as soon as their own space is made, and a denied folder
action lands in Activity after the attempt it denied. - Mastra's mark is its current one.
- The OpenBot image reports the release it is. A release candidate's image said the release before
it, so its pages sentOpenBot-Client: web/<previous>and anOPENBOT_MINIMUM_CLIENTSnaming
this release refused its own web app.openbot releaseprints it. - A step a rule refused, or one that failed, shows its whole reason, wrapped, where it was cut off.