MetaGPT

About

MetaGPT was built on the observation that LLMs produce incoherent software when asked for code directly, but produce far better results when forced through the intermediate artifacts a real team would create. A one-line requirement passes through product-manager, architect, and project-manager roles that emit user stories, competitive analysis, data structures, and API specifications before the engineer role writes code, with the pipeline exposed both as a CLI command and as a Python generate_repo call returning a ProjectRepo object. A DataInterpreter role extends the same machinery to data-analysis tasks. The project moved from the geekan personal account to the FoundationAgents organization and remains under MIT with a config2.yaml supporting OpenAI, Azure, Ollama, and Groq endpoints. Academic adoption (the ICLR 2024 paper and AFlow, an ICLR 2025 oral) keeps it a research substrate, while the commercial descendant MGX at mgx.dev carries the productized variant; teams use both to generate complete small projects from one-line requirements and as a multi-agent research substrate.