What it does
MetaGPT treats a small software company as a team of LLM agents. You give it a one-line requirement and it produces user stories, competitive analysis, requirements, data structures, APIs and other documents. Internally it includes product manager, architect, project manager and engineer roles that work through standard operating procedures, summed up by the project's idea that code is the SOP applied to a team. It installs with pip and needs Python 3.9 or later but below 3.12, plus Node and pnpm before real use. An example configuration file sets the LLM provider type (the example shows OpenAI, Azure, Ollama or Groq), the model name, a base URL and an API key. You can run it from the command line, for example with a prompt to create a 2048 game, which writes a repository to a workspace folder, or call it as a Python library. The repository also includes a Data Interpreter role for tasks such as data analysis with a plot, and lists use cases including Debate, Researcher and Receipt Assistant. It links to an online document, tutorials, a hosted demo on Hugging Face and a Discord community.
Capabilities
- Turns a one-line requirement into user stories, requirements, data structures, APIs and documents
- Role-based agent team: product manager, architect, project manager and engineer
- Run from the command line to generate a project repository, or import as a Python library
- LLM provider, model and API key set in a configuration file
- Data Interpreter role for data analysis tasks
When to use it
- You want to experiment with several LLM roles cooperating on a software task
- You want a starting set of design documents and a repository generated from a short brief
- You want to build your own agents using a role and SOP approach in Python
