
ChatDev
A graph of agents, scripts and human checkpoints, drawn in the browser or written as YAML
What is ChatDev?
ChatDev 2.0 builds a multi-agent system as a graph you draw in the browser or write as YAML: agent, python, human and subgraph nodes joined by edges that can branch, loop and fan out. You start a run from the Launch tab, watch each node move from pending to running to success or failed, answer a human node mid-run, and the same file runs from Python through the chatdev package on PyPI. The virtual software company from the 2023 paper is still here, rebuilt as one of the shipped workflows with nine agents from Chief Executive Officer to Software Test Engineer, but the 1.x code behind that paper now sits on a separate branch. The platform itself is young, and its own guides already disagree with the code on field names, so expect to learn it from the sample YAML files rather than the prose.
What can you do with ChatDev?
- Draw the graph, or write the YAML — make dev starts the backend and the web console together, and the console at localhost:5173 gives you a canvas of nodes and edges that exports back to YAML. The same workflows are plain files under yaml_instance/ — about forty of them ship, from chart building to deep research — and make validate-yamls checks every one for schema errors before a launch.
- Nodes that are not agents — A human node holds the run until someone answers in the console, attachments included. A python node takes the code block out of the message it just received and runs it in a shared code_workspace/. A subgraph node embeds another workflow by path or inline, and loop_counter and loop_timer stay silent until a cycle has run its rounds or spent its time, then release the edge that ends it.
- A different model on every node — Each agent node carries its own provider, base_url and api_key plus name, which is the model — the field the project's own guides get wrong in places. So a cheap model can draft and a stronger one review, and pointing base_url at Ollama or LM Studio keeps the run on your own machine.
- Fan a step out across an edge — Put dynamic on an edge and the target node expands into parallel copies: map splits the messages passing through and runs one copy per piece, tree does the same then merges results in groups until one is left. Splitting is per message, by regular expression or by JSON path, concurrency defaults to ten, and the docs warn this multiplies API calls.
- Attach a memory to an agent — simple keeps searchable conversation history, optionally saved to a JSON file. file builds a read-only knowledge base by chunking and indexing your own documents, rebuilt when they change. blackboard is a recency-ordered log with no search at all, and mem0 hands storage to Mem0's hosted service.
- Turn your own code into a tool — Drop a type-hinted function into functions/function_calling/ and its schema is built when the backend starts; the docs add a template-export command so the console's dropdowns list it too. Or attach an MCP server, either over HTTP or as a local process on stdio, which is how the Blender examples reach blender-mcp.
Before you choose ChatDev
- The virtual software company from the 2023 paper is now a sample workflow, not the product: the 1.x code behind it sits on the chatdev1.0 branch, where only news text has changed since December 2024.
- A python node runs the code block from its incoming message as a plain subprocess on the backend host, with a 60-second default timeout and no sandbox, so isolation means running the backend in the supplied container.
Star history
17 Aug to 30 Oct · +612
Frequently asked questions
Is ChatDev free for commercial use?
ChatDev is released under the Apache-2.0 licence — OSI-approved open source, which permits commercial use.
How can ChatDev be deployed?
ChatDev is available as Self-hosted / Runs locally.