Open Duet: Open-Source AI Coworkers You Run Yourself

Open Duet is an open-source, self-hosted workspace for AI coworkers: named agents that each have a job, instructions, memory, their own workspace and isolated computer, a schedule, and a human approval step before they do anything risky. You chat with them in a browser, a desktop app or Telegram, hand them recurring routines, and keep their data and credentials on your own machine or server. It is an open alternative to Claude Duet and to other paid AI agent products such as OpenAI Dots, Meta Muse, Grok Bot, Instinct, Manus Cue, Claude Cowork, and ChatGPT agent, and is not yet a feature-for-feature replacement for them.

An agent can be anything you would hand work to: a researcher, a support triager, a reporting assistant, a developer. Coding is one thing agents do, not what Open Duet is. Agents run on an engine you already have signed in (Claude, Codex, any Agent Client Protocol CLI, or an OpenAI-compatible endpoint). Teams of agents that delegate to each other, rooms, and connectors to other apps are planned, not built; see docs/ROADMAP.md.

▶ Watch the Open Duet demo on YouTube

Status: Active development. See Known limitations before running it for others.

Contents

• What it does
• Why Open Duet
• Quick start
• Using Open Duet
• Telegram and Boat sandboxes (optional)
• Architecture
• Project layout
• Known limitations
• Contributing
• License

What it does

• Named agents with a title, instructions, engine and model, workspace folder, approval mode and notes; export and import them as portable files.
• Memory: each agent keeps notes (MEMORY.md plus topic notes) that load into its conversations, with recall across conversations and undo.
• Routines: cron, weekday, interval, one-off and webhook schedules with timezones, run history, missed-run tracking, restart recovery and an in-app notification feed (see docs/ROUTINES.md).
• Approvals you control: ask-first, accept-edits and plan modes, saved rules, a decision log and a safety floor; risky actions wait for your Allow or Deny.
• Their own computer: give an agent an isolated Docker container so its commands and file edits run there, never on your machine (experimental, see docs/COMPUTERS.md).
• Durable conversations with streaming answers, a tool-call timeline, search, attachments, Stop, and resume after a restart.
• Several engines and limit-aware fallback: Claude, Codex, ACP agents and OpenAI-compatible endpoints; usage limits are recognised and a turn can fall back to another engine (see docs/ENGINES.md).
• Run it where you want: a desktop app from a source checkout or unsigned installers (experimental); or on a server you own, behind HTTPS, reachable from a browser or phone (single owner; experimental): python -m app.server, a Docker image and compose file, nginx and Caddy overlays, doctor, backup and restore. See docs/SELF-HOSTING.md.
• Channels: the web UI, the desktop app and a Telegram bot share one API.
• Your data stays yours: credentials are stored encrypted, in PostgreSQL (default) or SQLite (experimental, no Docker).
• Boat sandboxes (optional): the original Telegram and web task flow in hosted sandboxes, with scheduled prompts and GitHub push.

Why Open Duet

Paid AI coworker products keep your agents, memory and credentials on someone else's servers. Open Duet gives individuals and small teams the same idea in a form they can inspect and run themselves, with visible approval steps, local control over the app and its data, and no lock-in to one model provider. Read Known limitations before running it for others.

Quick start

Pick how you want to run it:

An engine is the model program your agents run on. You need at least one ready: a signed-in Claude Code or Codex CLI, an Agent Client Protocol CLI, or an OpenAI-compatible endpoint. Settings → Engines shows what is ready and why not.

Desktop app

git clone https://github.com/Anil-matcha/open-duet
cd open-duet/desktop
npm run app

npm run app sets up the first time, builds the page when needed, and opens Open Duet. It needs no .env, keeps its own encryption key in the operating system's secret storage, and uses a local SQLite database. On first start a setup flow in Settings detects your engines, asks you to pick a workspace folder, explains the approval modes and can create a starter agent. Details and troubleshooting: docs/SETUP.md. Tagged releases also build unsigned installers; they are unverified on a clean machine apart from a Windows smoke test, and there is no auto-update.

On a server

One owner, one data folder, HTTPS through nginx or Caddy, and engines signed in on the server. python -m app.server, a Docker image and compose file, doctor, backup and restore are included. Follow docs/SELF-HOSTING.md, and read what has and has not been checked there before relying on it.

Using Open Duet

A first session in the web app or the desktop app (the Chat page):

1. Sign in. The first start asks you to set the owner passphrase (docs/AUTH.md). Pair a phone or another browser from Settings → Devices.
2. Check your engine. Settings → Engines lists what is ready. Without a ready engine nothing can answer.
3. Create an agent. Chat → Agents → New agent. Give it a name, a title, instructions (its job, tone and rules), an engine and model, a workspace folder and an approval mode. Agents can be exported to a file and imported again.
4. Talk to it. Start a chat with the agent. Answers stream in, every tool call is a step you can open, and Stop works at any time. Conversations survive a restart and are searchable.
5. Approve what matters. A risky action shows a card with exactly what will run, and does not run until you press Allow. "Always allow this command" saves a rule you can review and undo in the Approvals panel. See Approval modes.
6. Let it remember. Each agent keeps notes (MEMORY.md and topic notes) that load into its conversations. Edit them under Agents → Notes. See docs/MEMORY.md.
7. Give it a routine. Chat → Routines runs a prompt as a new conversation of an agent on a schedule, by webhook or on demand, with run history, missed-run tracking and notifications. A routine never approves by itself. See docs/ROUTINES.md.
8. Give it a computer (optional). With Docker installed, Agents → Computer gives an agent its own isolated container so its commands and file edits never run on your machine. See docs/COMPUTERS.md.

Agents need LOCAL_EXECUTION_ENABLED=true on a server or development setup (the desktop app sets it itself). File edits and commands also need LOCAL_TOOLS_ENABLED=true, and every command still asks you first.

Telegram and Boat sandboxes (optional)

Open Duet started as a Telegram bot and web page that run tasks in hosted Boat sandboxes. That flow still works, is independent of agents, and needs the Boat key and Telegram setup described in Development setup.

Development setup, Telegram and Boat (optional)

Prerequisites

• uv for Python dependency management
• Node.js (20.9+) and npm, for the web UI (required by Next.js 16)
• Docker (for Postgres; not needed with the experimental SQLite option, see DATABASE_BACKEND below)
• Bash (Git Bash on Windows works fine) to use scripts/dev.sh
• Optionally, a Boat API key for sandbox creation. If it belongs to a personal account that also has an org/team on a paid plan, you also need that org's id (see BOAT_ORG_ID below) — otherwise sandbox creation bills the personal account and returns 402 Payment Required.
• A Telegram bot token from @BotFather (only needed if you want to run the bot, or let people link the web UI to Telegram — the web UI's guest mode works without it)
• A public URL the sandbox can reach to call back into your backend for permission decisions (see Permission prompts (Boat)) — in local dev this means tunneling your backend with ngrok or similar; not needed to run the app, only for any task that triggers a risky action (writing files, shell commands, etc.) to actually get a decision instead of hanging until it times out.

Configure the environment

scripts/dev.sh creates backend/.env from the example on first run, saves a private encryption key and webhook token, and preserves existing secrets on later runs. It also selects a trusted CA bundle for outbound HTTPS. Set SSL_CERT_FILE in your shell or in backend/.env to use a custom PEM bundle; certificate and hostname verification remain enabled. When Boat is configured, the script checks its TLS certificate before starting services, without sending credentials or creating a sandbox. You can start the API and web UI without a Boat key; sandbox features then remain unavailable.

For manual setup, copy the example yourself:

cp backend/.env.example backend/.env

Fill in backend/.env:

The web UI (frontend/) has its own .env.example — scripts/dev.sh copies it to .env.local automatically on first run. Copy it yourself if you're running the frontend separately, and point NEXT_PUBLIC_API_BASE_URL at the backend if you're not using the defaults.

Start the app

./scripts/dev.sh

From the repo root, this prepares local configuration, starts Postgres via Docker when PostgreSQL is selected, waits for it to be healthy, installs backend and frontend dependencies, runs migrations, and starts the FastAPI app (http://127.0.0.1:8000, docs at /docs), the Telegram bot when configured, and the web UI (http://127.0.0.1:3000). Without a Boat key, the API and web UI still start, while sandbox tasks stay unavailable. Ctrl+C stops everything it started (Postgres keeps running — docker compose -f backend/docker-compose.yml down to stop it too).

Skip pieces you don't need:

./scripts/dev.sh --no-bot        # API + web UI only
./scripts/dev.sh --no-frontend   # API + bot only

Run the pieces manually

cd backend
docker compose up -d                      # Postgres
uv sync                                   # install deps
uv run alembic upgrade head               # migrations
uv run fastapi dev app/main.py            # API (terminal 1)
uv run python -m app.telegram_bot         # bot (terminal 2)

cd ../frontend
npm install && npm run dev                # web UI (terminal 3)

Telegram and web task flow (Boat)

The original flow: tasks run in hosted Boat sandboxes, started from Telegram or the web page. Agents, routines and the Chat page below do not need it.

The Telegram bot walks a user through:

1. /connect — creates a sandbox and starts an OAuth login for the agent provider (currently Claude) inside it.
2. Pasting the OAuth code — the bot reads the resulting credential file out of the sandbox, encrypts it, and stores it.
3. /connect_github (optional) — links the user's own GitHub account via device-flow login, so tasks that git push do so as that user rather than a shared/anonymous identity. No token is ever typed into the chat: the bot posts a one-time code and a github.com/login/device link, the user approves it in their browser, and the bot picks up the result on its own.
4. Any plain text message — runs it as a task in the user's sandbox and replies with the result once it finishes. If the user has connected GitHub, that credential is restored into the sandbox before the task runs too.
5. /schedule — repeats a prompt on a cron schedule (hourly/daily/weekdays/ weekly, always in UTC) instead of running it once. /schedules lists them, /pause//resume//unschedule manage one. Each run is reported back in the chat — the result on success, the error on failure — and a schedule that can't succeed (e.g. an expired login) is paused automatically instead of retrying forever. See Scheduling tasks below.

The web UI covers the same ground from a browser:

1. Get started — either continue as a guest (an id the server makes, no Telegram needed) or link Telegram: the page generates a one-time code, you send /link <code> to the bot, and the server links your signed-in session to your real Telegram user id instead. (Sign in first: the web UI asks for the owner passphrase, see docs/AUTH.md.)
2. Connect Claude — same OAuth flow as /connect, driven by REST endpoints instead of chat messages (get-or-create the sandbox, start login, paste the code back into the page).
3. Connect GitHub (optional) — same device-flow login as /connect_github: the page shows a one-time code and a github.com/login/device link, and picks up the result on its own once you approve it — no code to paste back.
4. Submit a task — a prompt box that submits to the same POST /api/v1/tasks endpoint, polls for the result, and lists past tasks for that user. Finished tasks can be replied to in place to continue that same Claude conversation, instead of starting a new unrelated one. Checking "Repeat this on a schedule" turns the same prompt into a recurring schedule instead (POST /api/v1/schedules) — pick a frequency (hourly/daily/weekdays/weekly) and a time, in the browser's own timezone. Schedules are listed below the composer with pause/resume/delete controls. See Scheduling tasks.

In Telegram:

1. /start
2. /connect — creates a sandbox (can take a bit) and replies with a login URL. Open it, complete the OAuth flow, and paste the code back into the chat.
3. /connect_github (optional) — reply has a code and a link; open the link, enter the code, approve it. No need to send anything back in the chat.
4. Send any message, e.g. create a webpage that says hello — the bot runs it in your sandbox and replies with the result (or the error).
5. /schedule daily 08:00 generate my report — repeats that prompt every day at 08:00 UTC. /schedules shows it; /pause 1, /resume 1, /unschedule 1 manage it by the id the bot gave you. The API process must be running for it to actually fire — see Scheduling tasks.

In the web UI (http://localhost:3000), after setting the owner passphrase on first start (or signing in):

1. Get started — click "Continue without Telegram" for a standalone guest session, or "Link Telegram instead" and send /link <code> to the bot (the page picks this up automatically).
2. Connect Claude — click "Start login", open the link, and paste the code back into the page.
3. Connect GitHub (optional) — click "Connect GitHub", open the link, enter the code shown. No need to come back to the page — it updates once you approve it.
4. Submit a task — type a prompt and submit; the page polls for the result and lists past tasks below. Click a finished task to expand it and reply in that same conversation.
5. Check "Repeat this on a schedule", pick Daily and a time a couple of minutes out, and click "Create schedule". It appears in the schedules list below with its next run time; pause/resume/delete it from there.

Optional web search (Boat tasks)

Set PARALLEL_SEARCH_ENABLED=true in backend/.env and restart the backend and bot processes to make Parallel Search MCP available to new Claude task launches, including replies and scheduled runs. It connects over HTTP to https://search.parallel.ai/mcp and provides web_search and web_fetch without a Parallel account or API key. Anonymous search uses Fast mode and is intended for light use, with rate limits.

For example, submit: “Use Parallel to find the latest Python release, then fetch its release page and summarize the changes.” Approve the MCP tool calls through the existing permission prompts when asked. Claude still needs its usual connected account. The setting is off by default and leaves saved MCP configuration in place.

Scheduling tasks (Boat)

A schedule (schedule table) is a saved prompt plus a cron expression and timezone. Three background loops in the API process — not the bot process — drive it (app/core/background.py, app/services/ schedule_service.py, all on a 30s interval):

• poll_due_schedules — claims schedules whose next_run_at has passed and creates a Tasks row for each occurrence (FOR UPDATE SKIP LOCKED, safe with more than one API instance), then starts it.
• sync_active_tasks — finishes tasks that have no one polling them, the way a scheduled run does (nobody's watching a GET /tasks/{id} for it). Also recovers a task orphaned by a crash mid-launch.
• report_finished_runs — for each finished scheduled run, reported exactly once: sends the result on success, the error on failure, and pauses the schedule if it can't succeed on its own (an expired login, a missing credential) or after 3 failures in a row. Delivered over Telegram, to users identified by a numeric (Telegram-linked) user_id only — a web-only guest sees the outcome in the schedules list instead.

Because these loops live in the API process, the API must be running for a schedule to actually fire — creating one from Telegram doesn't require the bot process to also run the schedule (it only writes the row), but nothing fires without uv run fastapi dev app/main.py (or ./scripts/ dev.sh) up.

A schedule created from the web UI uses the browser's own timezone, captured at creation time. One created from Telegram is always UTC, since Telegram gives the bot no way to read the user's timezone — the bot's replies always spell out times in UTC.

Permission prompts (Boat sandboxes)

Every risky action the AI attempts inside a sandbox — writing a file, running a shell command, a git push, etc. — is intercepted by a PermissionRequest hook before it runs, and checked against a per-user rule table (permission_rule). No matching rule means the task pauses (status waiting_approval) and you're asked, in whichever channel you're using:

• Telegram: a message with Allow once / Always allow / Deny buttons.
• Web UI: the same three choices appear inline on the task once it's expanded, no page navigation needed.

"Always allow" writes a standing rule so the same action doesn't ask again; "Allow once"/"Deny" apply only to that one attempt. An unanswered prompt times out (110s) to a safe default of deny rather than hanging forever.

Shell (Bash) commands get a further pass: a small classifier (app/services/github_classifier.py) recognizes git subcommands and scopes the rule to the specific operation (git status read-only, git commit a local change, git push a publish) instead of treating every shell command the same — so "always allow committing" doesn't also silently allow pushing. Everything else run via Bash, and any tool without its own classifier, falls back to one coarse bucket per tool.

Extending classification to other connectors (beyond git) is still open; contributions welcome.

Architecture

Telegram bot ─┐                         ┌── PostgreSQL (default) or SQLite
Next.js web UI├── FastAPI backend ──────┤
Desktop app ──┘      │                  └── encrypted credentials
 (Electron,          ├── Boat sandboxes (optional): tasks, schedules, GitHub push
  local Claude)      ├── permission prompts and rules (git command classifier)
                     ├── task and schedule services (polling loops in the API process)
                     └── threads: local `claude` CLI turns, events stored then
                         streamed over SSE, approvals, resume after restart

Boat sandboxes are optional. Without a Boat key the app is chat-only unless local tools are enabled, and schedules, Telegram tasks and GitHub push are unavailable. The desktop app runs local Claude only. See docs/ARCHITECTURE.md for the target design.

Project layout

• backend/ — FastAPI app, Telegram bot, sandbox/auth/task services, Alembic migrations. See backend/README.md for backend-specific details (REST endpoints, env vars, DB models).
• frontend/ — Next.js web UI. See frontend/README.md for frontend-specific notes.
• desktop/ — Electron desktop app (local Claude, SQLite, no .env).
• scripts/dev.sh — one-command dev environment for the backend, bot and web UI.
• .github/workflows/ — CI (backend on SQLite and PostgreSQL, frontend lint/build, desktop tests) and the tag-triggered installer build.

Local Claude (experimental)

Optionally run prompts through your own signed-in claude CLI on the machine running the backend, with no Boat sandbox. It is off by default and only usable from that machine. The answer streams into the page as it is written. It is chat-only unless you set LOCAL_TOOLS_ENABLED=true, which lets Claude read and write files in its workspace folder and run commands, each call waiting for your Allow or Deny in the page (an allowed command runs as you, with no sandbox). See docs/SETUP.md.

A separate set of endpoints, /api/v1/threads, keeps conversations in the database so they continue and survive a restart. The Chat page (/chat, linked as "Open chat" on the home page) is their interface. See docs/SETUP.md.

Desktop app (from source, experimental)

An Electron window around the local backend, for local Claude only: from the desktop/ folder, npm run app (or the steps one at a time: npm run setup, npm run build, npm run start). It needs no .env file: it makes and keeps its own encryption key in the operating system's secret storage, creates its SQLite database, and only its own window can use the backend. npm run app does the setup and build on the first run and then opens it. Tagged releases build unsigned installers (Windows NSIS x64, macOS arm64 DMG) with GitHub Actions; neither has been verified on a clean machine apart from a Windows smoke test, and there is no auto-update. See docs/SETUP.md.

Known limitations

• Sandboxes are never automatically stopped/deleted — repeated /connect testing will accumulate them on the Boat account.
• Link codes expire after 10 minutes; there's no rate limiting on generating them.
• Authentication is single-owner: one passphrase per install, plus devices paired with one-time codes (docs/AUTH.md). There are no separate user accounts; every signed-in browser or paired device acts as the owner (a chat device only within conversations), and the user_id older routes take is not checked against the session. Sign-in lockouts are kept in memory per process. Use HTTPS beyond a trusted private network.
• Automated tests cover the permission engine and git classifier, schedules, local Claude (runner, stream, approvals, access policy), threads (store, runtime, stream, API), settings, the SQLite and PostgreSQL schemas and migrations, the desktop entry and launch token, and some task, sandbox and Telegram paths (backend/tests/, desktop/test/). Boat-sandbox task and sandbox services have only partial coverage, and there are no frontend tests. The GitHub Actions workflows have not yet run on GitHub, so treat them as untested until the first run.
• Engines are Claude, Codex, ACP agents, OpenAI-compatible endpoints and an OpenRouter preset of them; other engine presets are planned. Named agents and memory exist, but agents cannot delegate to each other yet, there are no rooms or teams, and connectors beyond an optional search MCP are not built.
• Desktop installers are unsigned; macOS and Linux have not been run.
• The git command classifier only covers the verbs it explicitly knows about (see github_classifier.py); an unrecognized verb defaults to the safer "mutation" bucket rather than being assumed read-only.
• Boat schedules (/schedules, Telegram /schedule) poll on a 30s interval, skip an occurrence missed by more than 15 minutes (recorded in last_error), and notify only Telegram-linked users (a numeric user_id) when TELEGRAM_BOT_TOKEN is set. Local routines (docs/ROUTINES.md) do not need Boat: they poll every 5s, record missed and skipped occurrences, notify in the app (and optionally one Telegram chat via ROUTINE_TELEGRAM_CHAT_ID), but only run while the backend is running, do not retry failed or interrupted runs, and do not de-duplicate webhook deliveries.
• Durable threads (/api/v1/threads) and the Chat page are tested on SQLite only, and store transcripts (including what Claude read, wrote and ran) in the database. A conversation can be deleted from there (not a secure erase), but the Claude CLI's own copy in your .claude folder is never touched.

Contributing

Issues and pull requests are welcome. Keep the README aligned with the backend and frontend behavior, and include focused tests when changing services or API behavior.

License

Open Duet is licensed under the Apache License 2.0.