MetaGPT

MetaGPT simulates a software company: hand it a one-line requirement and role-based agents acting as product manager, architect, and engineers follow structured operating procedures to produce user stories, competitive analysis, requirement docs, data structures, API designs, and working code. The framework popularized the assembly-line view of multi-agent collaboration, with the original paper accepted at ICLR 2024 and the follow-up AFlow work on automated agentic workflow generation taking an ICLR 2025 oral in the top 1.8 percent of submissions. A Data Interpreter role handles code-writing and analysis tasks, and the project works as both a CLI, where metagpt plus a prompt scaffolds a project, and a Python library. Installation is pip install metagpt on Python 3.9 to 3.11, with Node.js and pnpm required, Docker images available, and an LLM API key configured in YAML. The FoundationAgents org maintains it under the MIT license at around 69.6k stars, documents it at docs.deepwisdom.ai, and builds the commercial MGX agent development team product on the same research lineage.