ChatDev

ChatDev is an open-source platform where teams of AI agents build software together. The classic version runs a virtual software company: CEO, CTO, programmer, and tester agents carry a project from design to documented code. Version 2.0, called DevAll, is a zero-code orchestration platform where you define agents, workflows, and tasks through config files. The project comes from OpenBMB, is Apache-2.0 licensed, and has over 34,000 GitHub stars. A research paper describes the method.

Key Features

1. Virtual software company

Role-based agents split the classic software process: design, coding, testing, and docs. Chat records show every decision.

2. DevAll zero-code orchestration

Version 2.0 defines agents and tasks in config, no code needed. It also handles data viz, 3D generation, and deep research.

3. Human review mode

You can step in as a reviewer at any phase. Agents pause for your input before moving on.

4. Git mode

Projects get real version control as agents work. You can track, branch, and roll back agent-written code.

5. Art mode

A designer agent generates images and visual assets for the project. Design and code stay in one pipeline.

6. Any LLM provider

Point it at OpenAI or another provider through a simple config file. You are not tied to one model company.

Frequently Asked Questions (FAQ)

How do I install ChatDev?

Set up the backend with uv sync and the frontend with npm install, then add your API key to a config file. Docker Compose is also available.

What is the difference between 1.0 and 2.0?

Version 1.0 is the virtual software company with fixed roles. Version 2.0 is a general orchestration platform you configure for any multi-agent task.

What does ChatDev cost to run?

The code costs nothing: it is Apache-2.0 licensed and free to use. You only pay your LLM provider for the model calls agents make.

Where can I read the research?

The method is published as a paper on arXiv. It explains how chat chains let agents collaborate on software tasks.

Conclusion

ChatDev makes multi-agent software building concrete instead of theoretical. Watching designer, coder, and tester agents trade messages is both fun and instructive. Version 2.0 widens it into a general orchestration tool. If you want to study agent teamwork on real code, this is one of the best open labs available. It remains a top open lab for this.