Open Duet: Raise a team of AI colleagues on your own computer or server, with a name, a position, and a memory, and you need to nod before you startOpen Duet：在自己电脑或服务器上养一支 AI 同事团队，有名字、有岗位、有记忆，动手前还要你点头

Each agent has independent instructions, models, workspaces, memories, and approval settings: one to do research, one to write protocols, one to run support, and one to write code, without interfering with each other.

You can set up a lead and a number of specialists to let the lead decide for themselves which piece of work to give to whom; their work runs in their respective conversations and returns to the lead to summarize it. It is also possible to put several agents in the same room to discuss together.

Typical usage: build a lead, a researcher, an editor, share a project folder, and let the lead save a weekly report from the notes and reports - once satisfied, turn it into a weekly automatic running routine.

Memory: Agents can take notes, recall relevant notes and historical conversations in follow-up conversations. Notebooks can be viewed or changed, not black boxes.

Routines: triggered by schedule, by webhook, or manually “run now”; run on the back end so the back end is always on; the key is that it will never approve the action itself and will stop waiting for you when you need to nod.

Three levels of approval: Ask first, Accept edits, Plan only. The tool call record and approval decision can be reviewed in the interface.

Tasks and calendars: Put your work on a Kanban board, assign it to an agent or team, and see both deadlines and Routine schedules in the calendar; also count token usage and spending.

The engine can be connected to: Claude Code, Codex CLI, Gemini CLI, Cursor, or any command-line proxy that takes the ACP (Agent Client Protocol); it can also be connected to OpenRouter, or OpenAI compatible interfaces (local inference such as Ollama, LM Studio).

It does not itself contain model credits, nor does it sell subscriptions - you bring your own account, CLI login, or API key. This is why it is cheap, and it is also the premise of its "data does not come from its own machine".

Different engines can do different things: such as image input, agent computer (Docker isolation environment) is not supported by every engine.

Desktop side (source code running, the most convenient): requires Node.js 22 +, uv, and a CLI engine that is installed and logged into the machine. Three commands - git clone repository, cd open-duet/desktop, npm run app. Start the wizard for the first time to check the engine, select the workspace, select the approval mode, and build a starting agent.

Own server (computer shutdown also continues to run): Linux server + domain name with HTTPS + Docker (or Node.js and uv) + an engine on the server. Positioning is a single owner, paired with their own device, not a multi-user organization workspace.

Development checkout: If you want to change the code or connect to PostgreSQL, go to this one, see the repository docs/SETUP.md.

Official v0.4.0 provides Windows x64 installation package (133MB) and macOS arm64 DMG (151MB), both of which exceed the 25MB limit for single files on this site, so the download button is given the official source code package (zip, 1.78 MB).

Press the above three commands to run from the source code to the desktop; if you really want to install the package, go to the warehouse Releases page to download it yourself, pay attention to the official mark that these two installation packages are still unsigned and belong to the experimental state.

The project is still in active development, team delegation, and server mode are all new features; before using it to handle important work, read the Known limitations in the warehouse.

The model call fee is charged by your own service provider, Open Duet does not provide models. This article is compiled from public information and accompanied by supporting resources; functions and prices are subject to the official page. Resources are for learning communication only, please follow the Apache-2.0 license.