OpenDuet

OpenDuet is CopilotKit's self-hosted template for a proactive agent. It starts the day with a short brief, keeps working on a computer of its own, and pauses before it sends, pays, signs, or deletes. Checked 3 November 2026 against the GitHub API, the README, the tenant YAML, triage.ts, brief.ts, morning-brief.ts, telemetry.ts, gateway.ts, policy-store.ts, config.ts, copilot.ts, channels/routes.ts, chat-screen.tsx, the head of calls.ts, docs/telemetry.md, evals/triage/README.md, and the public page at copilotkit.ai/openduet. Not cloned and not started.

What you get

The GitHub description calls it an open-source template for a proactive AI agent that gets your work done and keeps you in the loop. The README badge says early development. What makes the checkout OpenDuet, rather than plain OpenBot, is the name, the bots, the channels, the model, and the skills in openbot/examples/openduet. The platform under openbot/ is a copy of OpenBot committed in this repo. A request for .gitmodules returned 404, so the copy is not a submodule. That copy was not diffed against CopilotKit/openbot. The tenant folder is not the only OpenDuet code. wiring.ts says index.ts is OpenBot's, and every OpenDuet line added there can collide with an OpenBot update. openbot/server/src/openduet/ also holds tasks, pages, channels, replies, and offers. Those files were listed. Their bodies were not read, except brief.ts and triage.ts.

The README says OpenDuet is CopilotKit's open-source take on assistants that reach out first, such as Claude Duet, and that it is not affiliated with or endorsed by Anthropic.

The product page, fetched 3 November 2026, says early preview and MIT. It says you self-host with Docker and Bun. It says policy, approvals, the audit record, and saved logins stay in your PostgreSQL, while threads and memory stay in CopilotKit Intelligence. The page lists LangChain, Mastra, Google ADK, CrewAI, Microsoft Agent Framework, and AWS Strands. The README lists LangGraph, Mastra, CrewAI, Pydantic AI, Google ADK, or an agent written by hand. The tree contains agent-langgraph, agent-mastra, agent-crewai, agent-pydantic-ai, agent-adk, agent-microsoft, agent-strands, and other agent-* folders. Those folders were not opened.

How a turn runs

You talk in the web app on port 3010. Slack and Microsoft Teams are OpenBot's, per the README. .env.example says those two arrive through an OpenTag deployment, and that the old SLACK_* settings are refused at startup. That path was not exercised.

The server package depends on Hono ^4.10.0 and @copilotkit/runtime 1.77.0. It sends each turn to an agent over AG-UI. @ag-ui/client and @ag-ui/core are 1.0.1. openbot/examples/openduet/agents.yaml defines three bots.

openduet is type: built-in. Its prompt tells it to stay under 35 words in ordinary chat, to call the tool for send, pay, sign, or delete without asking first in the chat, and to put longer work in make_page or start_task. The pause the person sees is the approval gate, not a second question in the thread.

your-agent is type: remote-ag-ui. Its endpoint is ${MANAGED_AGENT_AG_UI_URL:-}. A blank endpoint leaves this bot out instead of failing startup. .env.example says scripts/start.sh points that URL at the LangGraph agent on port 4201 for a laptop, and that the proof of concept on port 4200 is a reference, not something to deploy.

picked-harness is kept so OpenBot's harness picker in scripts/start.sh still works.

Computer actions and connector calls take different paths.

The header of openbot/server/src/computer/gateway.ts says that file is the only way an action reaches a bot's computer. It resolves the element ref from the server's snapshot, asks policy, writes the audit row, and only then acts. Deny beats allow. An absent policy denies. A broken rule denies. The comment says a label supplied by the model is not what gets clicked. DEFAULT_ACTION_POLICY in policy-store.ts is mode enforce, with an empty deny list and allow: ["true"]. The comment says that allow is deliberate, so a new install can act and an administrator has a place to write the first restriction. The approval gate is what pauses a write. The buttons that map allow_once, allow_always, and deny onto Allow once, Always allow, and Refuse were not opened.

Connector tools go through callTool in openbot/server/src/plugins/plugin-store/calls.ts. A write that shareTargetOf treats as a message to other people asks the owner first, even when the general approvals switch is off. The rest of that file was not read.

openbot/worker/src/index.ts sets TICK_MS to 30 seconds and sweeps due routines in the same process, because a laptop has no CronJob. The handler for /internal/routines/run was not opened.

The README says each agent can have a container with Chromium, a /workspace volume, a browser profile, and a shell. Two switches are easy to mix up. The README says keep OPENBOT_ONE_COMPUTER_EACH=true for a computer per bot, and set it to false with an empty COMPUTER_SUPERVISOR_URL to share one. .env.example says an unset COMPUTER_SUPERVISOR_URL means every bot shares AGENT_COMPUTER_URL, and setting it makes the supervisor give each bot its own container. COMPUTER_RUNTIME=runsc uses gVisor when the host has it. Unset, the example says a computer is an ordinary container on the host kernel.

Mail triage and the morning brief

openbot/shared/morning-brief.ts exports MORNING_BRIEF_TOOL as morning_brief. The frame text is Give me my morning brief. It tells the model that nobody typed it, to call morning_brief once, and to add at most one short sentence after the card.

The browser does not send that text. chat-screen.tsx POSTs { agentId, date } to /api/channels/morning-brief. The comment there says the server writes the message. routes.ts ignores a brief field an older app might still send, and calls morningBriefMessage(). briefDate accepts YYYY-MM-DD only when the date is within one day of the server's UTC date. The first-day update sets morningBriefOn to that date. The comment says a later open on the same day does not also post a brief. first-day.ts was not read.

buildMorningBrief in openbot/server/src/openduet/brief.ts searches Gmail with in:inbox newer_than:1d unless since is set, and caps the fetch at BRIEF_MAIL_CAP, which is 50. It sorts eight messages at a time. The text it sorts is the search snippet, not the full message. A message that fails to sort is listed under needs you, with the reason "Could not be sorted, so it is here to be safe." Calendar events come from google-workspace/calendar_list_events. If Gmail or Calendar errors, the brief records that source under couldNotCheck. A comment in that file says four-at-a-time sorting of 50 messages took about 50 seconds against the fake mailbox. That timing was not re-run.

triage.ts has two stages. ruleVerdict can only return can_wait. The rules are bulk (a list or bulk header, and a Gmail Promotions or Social label), category (those two labels alone), auto-reply (Auto-Submitted: auto-replied, not auto-generated), and calendar-response (a subject that starts with accepted, declined, or tentatively accepted). Updates and Forums are left for the model. A noreply sender is not a rule. Anything the rules do not settle goes to the model through triage. decidedBy is rule:<id> or model:<model name>.

evals/triage/README.md says the set is 62 invented messages to Maya Okafor on the morning of Thursday 22 October 2026. 27 need her and 35 can wait. No real mail is used. The committed notes in that README are claude-sonnet-5-5 on all 62, with precision 100% and recall 100% for needs you, and 12 decided by rules. gpt-5.6-terra is on the earlier 60, with precision 96.2% and recall 100%. The same README says the set is small, the same person wrote the labels and the prompt, and the numbers are a regression check, not proof on a real inbox. The eval was not re-run.

The morning-brief skill tells the bot to call morning_brief and not to list the items again. The app has morning-brief-card.tsx and morning-brief.tsx. Those components were not read.

The tenant package

Copy openbot/examples/openduet and point TENANT_PACKAGE_DIR at the copy. The path is resolved from openbot/server. .env.example sets TENANT_PACKAGE_DIR=../examples/openduet.

The comment in skills.yaml says naming a tool there grants nothing. The offer is intersected with the bot's grants. reply-to-an-email tells the bot to call gmail_send_reply without asking in chat first, because the send itself pauses. research-public-web names @research/search and @research/fetch. The README says an administrator enables a connector from Parallel, Exa, or Firecrawl and grants those tools. openbot/docs/web-research.md was not opened.

What you need to run it

From the repository root, the README's quick start needs Docker, Bun 1.3 or newer, Node so npx can run the CopilotKit CLI, a model key, and a CopilotKit Intelligence project. Intelligence is required. runtimeCapabilities in openbot/server/src/config.ts throws CopilotKit Intelligence is required and is not configured and names whichever of INTELLIGENCE_API_URL, INTELLIGENCE_GATEWAY_WS_URL, and INTELLIGENCE_API_KEY is missing. docs/telemetry.md only mentions the key. Whether scripts/start.sh fills the two URLs was not traced.

git clone https://github.com/CopilotKit/OpenDuet.git
cd OpenDuet
cp .env.example .env
bun install
npx --yes copilotkit@latest login
npx --yes copilotkit@latest project select
bun openbot/scripts/setup-learning.ts
bash scripts/start.sh

Then open http://localhost:3010. scripts/start.sh links this repo's .env and execs openbot/scripts/start.sh. scripts/stop.sh stops the app, the worker, the API, the Docker services, and this install's bot computers. It does not delete the database volume, the workspace files, or the browser profile.

project select writes CPK_INTELLIGENCE_API_KEY and CPK_TELEMETRY_ID. The server reads INTELLIGENCE_API_KEY. setup-learning.ts copies the key and sets CPK_INTELLIGENCE_LEARNING_CONTAINER_ID=openbot. The README says the free Developer plan is enough. That plan was not signed into.

The OpenDuet bot's model in model.yaml is gpt-5.6-terra. runtimeModelForEnvironment in copilot.ts keeps the tenant model unless BOT_PROVIDER is anthropic, or BOT_MODEL is set and OPENAI_BASE_URL is also set. The built-in OpenAI fallback constant in that file is also gpt-5.6-terra. The tool loop stops at TOOL_STEPS, which is 8. For Anthropic, set BOT_PROVIDER=anthropic and ANTHROPIC_API_KEY. The model is BOT_MODEL if you set one, and claude-sonnet-4-5 if you do not. .env.example separately says the framework bot's BOT_MODEL defaults to gpt-5.5, and the proof-of-concept bot's AGENT_BOT_MODEL defaults to gpt-5.5. Those defaults were not traced through start.sh.

PORT and SERVER_PORT are both 3001 in the example. The comment says the server prefers PORT when both are set and refuses to start if they disagree. The app's Vite proxy reads SERVER_PORT. If 5432 is taken, set POSTGRES_PORT and change the port in DATABASE_URL. The example URL is postgres://openbot:openbot@localhost:5432/openbot. COMPOSE_PROJECT_NAME=openduet keeps Compose from sharing containers with a checkout whose compose file lives in a folder named openbot. Image tags in the example are openduet-server:latest and the other openduet-* names listed there.

The README says a hosted deploy guide and an OpenDuet image are not published. Docker and Helm docs under openbot/docs deploy plain OpenBot from OpenBot's image, not this template.

Root package.json is private, name openduet, version 0.1.0. openbot/package.json is private, name openbot, version 0.2.0, packageManager bun@1.4.3. There are no GitHub releases and no tags. openbot/CHANGELOG.md starts at Unreleased, then 0.2.0.

Telemetry and sign-in

openbot/server/src/telemetry.ts sets TELEMETRY_APP_TITLE to OpenDuet and sets COPILOTKIT_TELEMETRY_SAMPLE_RATE to 1 when that variable is empty. docs/telemetry.md calls 1 full sampling and says the SDK default is 5%. The title is fixed. A renamed fork is still counted as OpenDuet unless OPENBOT_ACCESSIBILITY_DISABLED=true. Events go to https://telemetry.copilotkit.ai/ingest. The doc names oss.runtime.instance_created, oss.runtime.copilot_request_created, and intelligence_signup. Intelligence sends the last one when an account is created. This repo does not. Opt out with COPILOTKIT_TELEMETRY_DISABLED=true or DO_NOT_TRACK=1, then restart. The tests replace fetch and do not prove the events arrive.

.env.example sets OPENBOT_SINGLE_USER=true. The comment says every visitor is an administrator while that line is set, and that with no sign-in provider the listen address is this machine only unless OPENBOT_HOST widens it. Delete the line and configure Google, Microsoft, or Okta before anyone else can reach it. With no provider and the line removed, the deployment refuses to start. The example KEY_ENCRYPTION_KEY is a public development value. The comment says production refuses to start with that key.

Composio is optional. Without COMPOSIO_API_KEY there is no brokered catalogue. The example says there is no screen that sets the key.

Gmail and Calendar need a Google OAuth client, then each person connects under Connections. OPENDUET_FAKE_GOOGLE_URL points those connectors at openbot/scripts/fake-google/server.ts and is ignored when NODE_ENV=production. The fake steps in the README were not run.

Repo snapshot

GitHub API on 3 November 2026. Created 2026-10-21T19:43:59Z. Last push 2026-10-31T21:26:41Z. updated_at 2026-11-02T20:02:16Z. Default branch main. The newest commit in the first page is f52288fff5, a README badge merge. Language TypeScript. License MIT, copyright 2026 CopilotKit. 1,847 stars, 233 forks, 11 open issues, 8 watchers. Topics empty. The API size field is 164,245 KB. Homepage field copilotkit.ai/openduet. https://www.copilotkit.ai/openduet returned 200. https://copilotkit.ai/openduet returned 308.

CopilotKit/openbot, read the same day and not diffed, had 9,885 stars, 1,211 forks, MIT, and a push at 2026-11-02T19:06:20Z.

Not checked

Not cloned, not installed, and not started. No Intelligence login, no model call, and no eval re-run. The password box, the approval policy defaults, pgvector, and the OpenTag path were not read past the README and .env.example. The README diagram names PostgreSQL and pgvector for policy, approvals, audit, and encrypted credentials. The schema was not opened to confirm the extension. The vendored openbot/ tree was not compared with CopilotKit/openbot. The Tauri app under openbot/desktop and the Expo app under openbot/mobile were not built. evals/reply-chips and evals/voice were not scored again.

See also: Agent Infrastructure And Platforms