Open Duet

Your own team of AI coworkers, running on your computer or a server you control.

Open Duet is an open-source workspace for creating named AI agents, giving each one a job and a place to work, and chatting with them over time. Let a lead agent hand work to specialists, keep useful context between conversations, or set a routine to run later. Choose how much freedom an agent has, and review the actions it takes.

You choose where the app runs and which engine gets your prompts. Open Duet does not include a model or a model subscription.

Project on GitHub · Desktop app · Web app (invite-only) · Watch the demo · Quick start · Run it on your server · Guides

Status: Active development. The desktop app, self-hosted server, agent computers, and newer team features are still evolving. See known limitations before relying on them for important work.

• Create a roster of agents. Give each one a name, instructions, model, workspace, memory, and approval settings. Use different agents for research, planning, writing, support, or development.
• Work as a team. Set a lead and specialists, let the lead delegate tasks, and follow the results in chat. You can also bring agents together in shared rooms. Team leads can use Claude or Codex. See teams and delegated work.
• Keep useful context. Agents can keep notes, recall relevant notes and earlier conversations, and search their own history. You can review and edit what they remember. See memory.
• Set work to run later. Create routines that run on a schedule, when another service triggers them, or when you press Run now. Follow each run and respond if it needs your approval. Routines do not need Boat. See routines.
• Stay in control. Choose Ask first, Accept edits, or Plan only. Review tool activity and approval decisions in the app. Agents can also be given an isolated Docker computer, a Boat cloud computer (desktop app only), or a local browser; computer and browser access are experimental and off until you enable them.
• Connect your apps. Browse and connect apps such as GitHub, Gmail and Google Calendar from the Apps page (through your own Composio key), then choose which Claude or Codex agents may use which connected accounts. Calls that change anything wait for your approval. See connected apps.
• Add tools from the MCP Marketplace. Browse a GitHub-backed catalog of community and tested remote MCP servers, store credentials separately, and grant an agent exact tools, with one-time approvals for calls. Source implementation only; live-service qualification remains.
• See what needs you. The Today view gathers pending approvals and questions, routine outcomes, engine failures and due tasks, and can show read-only Gmail and calendar items. Routines can report quiet results and draft replies for your review. Ask for a refreshed brief from chat. See everyday features.
• Reuse instructions. Save skills, which are text instructions you pick per message, and start new agents from templates. See skills and templates.
• Use it from your phone. Pair the phone app with your own server to chat, stop an answer and allow or deny a tool call. It is early and source-only; see the phone app guide.
• Keep track of work. Search conversations, attach text files and images, import audio for transcription (with your consent and credentials), browse or download workspace files, export your history, and review reported token use and cost.
• Plan tasks and deadlines. Keep work on a board, assign an agent or team, and review deadlines alongside routine occurrences in a week or month calendar. Start an agent from a card, schedule a one-time run, and review its conversation and attempt history. See tasks and calendar.

Create a lead agent and a couple of specialists, such as a researcher and an editor. Give them the same project workspace, and set the lead's engine to Claude or Codex. Then ask the lead to prepare a briefing from the notes and reports in that folder.

The lead can split the request into smaller tasks and hand them to the specialists. Their work runs in its own conversations; when the results are ready, they return to the lead, which can put the briefing together. You can open the chats to follow along, and approve actions when the approval settings require it. Once the task is useful, make it a weekly routine and review each run from the app.

Open Duet stores conversations, agent notes, routines, and settings in the local data folder or server database for the instance you run. Workspace files stay in the folders you assign to agents. API keys saved through Open Duet are encrypted there; command-line sign-ins remain with the engine installed on that machine.

When an agent answers, the selected engine receives the prompt and the context needed for that turn. That can include text you attach, relevant notes, and workspace content the engine reads. If the engine is a hosted service, that information is handled under that provider's terms. Open Duet does not include a hosted model or a model subscription.

Connect an engine you already use: Claude Code or Codex CLI, Gemini CLI, Google Gemini through Antigravity, Cursor, another command-line agent using the Agent Client Protocol (ACP), OpenRouter, or an OpenAI-compatible API such as Ollama or LM Studio. Available tools vary by engine; for example, image input and agent computers are not supported by every engine.

See engine setup and capabilities. You bring the provider account, CLI sign-in, or API key; model usage may cost money.

The desktop app runs from a source checkout. It needs Node.js 22 or newer, uv, and at least one engine available on the same computer. For a CLI engine, install it and sign in before opening Open Duet. You can also configure an API-based engine in Settings → Engines. You do not need Docker for ordinary chat; Docker is only needed for the optional isolated agent computers.

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

On first launch, the setup screen checks your engines, helps you choose a workspace and approval mode, and can create a starter agent. Then:

1. Open Settings → Engines and confirm that an engine is ready.
2. Create an agent with a name, a clear job, and the workspace it should use.
3. Start a chat. Choose Ask first if you want the agent to pause for approval before actions that need it.
4. Add teammates or a routine when you want to split up work or repeat it later.

As of 2026-10-30, the Releases page has a Windows x64 installer and macOS arm64 DMG for v0.7.0 (listed as a pre-release, so it is not under "latest"). The release includes the MCP Marketplace, cloud computers, and recent Today and memory fixes; see the changelog. Both installers are experimental and unsigned. The macOS installer omits native desktop preview; that preview remains available from a source checkout with the reviewed bundle staged. The app includes an assisted update flow that offers a SHA-256-verified installer for you to open; it does not install updates automatically. The release-to-release update path has not been verified end to end.

Routines run prompts on local agents on a schedule, by webhook, or on demand. They run in the backend, so the backend must stay running. A routine never approves an action by itself. See the routines guide.

The original Telegram bot and Boat sandbox task flow are also available as an optional, separate path. Those scheduled tasks use the API process too; see backend details and environment settings. Neither Telegram nor Boat is needed for the agent workspace.

• Open Duet is early and changing. The server setup is experimental and intended for one owner with paired devices; it is not a multi-user organization workspace.
• Team delegation is new; lead agents need to use Claude or Codex.
• The phone app is early: source only, tried on an Android emulator but not on iOS or a physical phone, with no push notifications and no store release.
• Docker computers and their browser tools are experimental and opt-in. The macOS native desktop preview is also experimental and available only in the desktop app.
• Desktop installers are unsigned and have limited platform verification. See desktop setup details.
• Open Duet does not provide model access. You are responsible for your model provider's account, terms, and usage costs.
• Connected apps are reached through Composio and tested against a fake Composio, not yet the real service. Only Claude and Codex agents can use them, and every change to an app waits for your approval. The MCP Marketplace, Boat cloud computers, quiet routine results and audio transcription have not had live-service or packaged-app qualification. The optional Telegram and Boat task flow remains separate from the agent workspace.

• First-run setup and approvals
• Engines and models
• Teams and delegated work
• Agent settings and chat management
• Memory and context
• Routines
• Tasks and calendar
• Connected apps
• Skills and templates
• Everyday features: search, attachments, audio, Today
• Agent computers, native desktop and browser tools
• Features, capability matrix and roadmap
• Changelog
• Self-hosting
• Phone app
• Contributing

Issues and pull requests are welcome. Start with the contributing guide.

To propose a third-party app for the in-app MCP Marketplace, add a Community manifest under marketplace/plugins/community/, regenerate marketplace/catalog.json, and submit both in a pull request. See the MCP Marketplace submission guide for the manifest requirements and validation steps.

Open Duet is licensed under the Apache License 2.0.