ChatDev

ChatDev 2.0 (DevAll) treats a multi-agent system as a graph declared in YAML. Nodes are LLM agents, Python steps, human checkpoints, nested subgraphs or small utility nodes; edges carry messages and can have keyword or other conditions. A FastAPI backend validates the file, creates a session and runs the graph, while a Vue web console and a Python SDK are the two ways to start runs.

Acyclic graphs run by topological sort, with nodes in the same layer executed concurrently. Loops are found with Tarjan's algorithm and wrapped as super nodes so the outer graph stays acyclic; loop-counter and loop-timer nodes bound iterations. Dynamic edges can split a message and run many copies of the target node.

Each run gets its own session folder under WareHouse/ holding attachments, a shared code_workspace/ and context snapshots. Longer-term memory is opt-in: stores declared under memory are attached to agent nodes with read/write rules. The older chat-chain design with CEO, CTO and programmer roles is kept on the chatdev1.0 branch.

A human node blocks the workflow and shows its description plus the current context in the Launch view; the person types a reply or uploads files, and the text flows to downstream nodes. Combined with a keyword edge condition, this gives review loops (for example, repeat until the reviewer types ACCEPT). Runs can be watched live in the web console. No per-tool-call approval mechanism is documented.

uv sync

# yaml_instance/release_notes.yaml - run it from the Launch tab after `make dev`
version: 0.4.0
vars:
  API_KEY: ${API_KEY}
graph:
  id: release_notes
  description: Draft release notes, then wait for a human to approve
  start: [Drafter]
  end: [Drafter]
  nodes:
    - id: Drafter
      type: agent
      config: {provider: openai, name: gpt-4o, api_key: "${API_KEY}", role: "Write concise release notes from the input."}
    - id: Approver
      type: human
      config: {description: "Type ACCEPT to finish, or give revision notes."}
  edges:
    - {from: Drafter, to: Approver}
    - {from: Approver, to: Drafter, condition: {type: keyword, config: {none: [ACCEPT]}}}

• uv sync is run inside a clone of the repository; the frontend also needs cd frontend && npm install.
• Requires Python 3.12+, Node.js 18+ and uv.
• Copy .env.example to .env and set API_KEY and BASE_URL; an undefined ${VAR} placeholder raises a ConfigError.
• If the web console cannot reach the backend, port 6400 may be taken; change both --port and VITE_API_BASE_URL.
• MCP tooling must use mcp_remote or mcp_local; the old type: mcp schema is no longer accepted.
• An open issue reports the Docker setup failing with sh: vite: not found.

The current main branch: a zero-code platform where multi-agent workflows are YAML graphs run through a web console or Python SDK. It replaced the ChatDev 1.0 virtual software company, which now lives on the chatdev1.0 branch.

Yes, as a client. Agent nodes can use tools from remote HTTP MCP servers (mcp_remote) or local stdio MCP processes (mcp_local). A2A and AG-UI support were not found.

Yes, the code is Apache-2.0. You pay your model provider. A 2023 README news item mentions a SaaS at chatdev.modelbest.cn, but no pricing for it is documented.

Yes. A human node pauses the run and asks for input in the Launch view; an edge condition can loop back until the person types a keyword such as ACCEPT.

The README documents an OpenClaw integration: after installing the chatdev skill, OpenClaw can invoke existing ChatDev agent teams or create new ones through the running ChatDev backend.