Graphic tutorial: build your own team of AI colleagues in half an hour, from setting up the environment to letting lead party and regularly running weekly newspapers图文教程：半小时搭起自己的 AI 同事团队，从装环境到让 lead 派活、定时跑周报

Node.js 22 or later; uv (Python package manager, used in scripts); and at least one available engine.

The most troublesome part of the engine is the Claude Code or Codex CLI installed on the machine and logged in - the command line can run normally, and Open Duet can take over. You can also use the Gemini CLI and Cursor to use the ACP protocol.

I don't want to install the CLI at all, just add an OpenAI compatible interface (Ollama, LM Studio) or OpenRouter key in Settings → Engines, the effect is the same, but the available tool capabilities will be less.

Three commands: git clone https://github.com/Anil-matcha/open-duet.git → cd open-duet/desktop → npm run app. For the first time, I will pretend to be dependent, patient, etc.

Launch the Setup Wizard for the first time: it will check that the engine is ready, let you select a workspace directory, select the default approval mode, and easily create a starting agent.

After entering the main screen, do one thing first - open Settings → Engines and confirm that an engine shows ready; this step is not passed, and all subsequent conversations will fail.

Note: Windows exe/macOS DMG is also officially available, but they are more than 25MB and are not signed. This tutorial has a more controlled route to the source code.

Click to create a new agent, and write down four things: the name (such as "research assistant"), the job command (what it is responsible for, what format it outputs, and what circumstances should stop asking), the workspace (the folder that can only be read and written), and the model (which engine).

Ask first for approval mode: It stops every time you need to move a file or run a command and waits for you to nod. The first few times you just took the opportunity to see what it did.

Then send a real task in the conversation, such as “Read the three materials in the workspace and list five questions worth following up.” Observe its tool call record and confirm that the behavior meets expectations before relaxing to Accept edits.

Build two more agents: one researcher (responsible for checking materials and organizing key points), one editor (responsible for completing manuscripts and unifying calibres), and three agents pointing to the same project work area.

Set the first agent to lead and the engine must be Claude or Codex (only these two are supported by the current team delegation).

Say to lead: "Let the researcher write a 200-word bulletin for each of these three materials, and you will compile it into a weekly report and submit it to the editor for a unified tone." It will disassemble tasks, dispatch, and receive results by itself; if you want to see the details, you can click on the corresponding sub-session.

Memory is useful here: Researchers write down conclusions that can be recalled in the next week's session; you can change the misremembered entries in the memory panel.

The weekly report process is running smoothly, just create a new one in Routines: prompt write "Generate the project weekly report in the format of last week, the material is taken from the workspace/weekly directory", the trigger method is selected every Monday morning, or let another system trigger it for a webhook.

Remember two hard rules: Routine runs on the back end, so the back-end process must always be on; it will never approve the action itself, it will stop waiting for you when it needs to be approved, and you can see the progress and pending items of each run in the app.

If you want the computer to shut down and continue to run, take the docs/SELF-HOSTING.md part to the Linux server (domain name + HTTPS + Docker or Node), position it for single use, device pairing, and do not use it as a multi-person organization platform.

Finally, take a look at the dosage page: the token cost is calculated by agent and session, and it is clear which agent is burning money.

The model call fee is charged by your service provider. The project is in the active development stage. Team delegation and server mode are new features. Please read the Known limitations of the warehouse before important work. Resources are for learning exchanges only and are licensed under the Apache-2.0 license.