Capability · Framework — agents

MetaGPT

MetaGPT applies a software-company metaphor to agent orchestration: a Product Manager agent writes a PRD, an Architect agent designs the system, Engineers implement, and QA reviews. Each role has its own prompts and SOPs (standard operating procedures) encoded in the framework — producing surprisingly structured codebases for well-scoped problems.

Framework facts

Category: agents
Language: Python
License: MIT
Repository: https://github.com/geekan/MetaGPT

Install

pip install --upgrade metagpt
metagpt --init-config   # fills ~/.metagpt/config2.yaml

Quickstart

import asyncio
from metagpt.software_company import generate_repo

async def main():
    repo = generate_repo('Create a 2048 game in Python with PyGame')
    print(repo)

asyncio.run(main())

Alternatives

CrewAI — lighter role-based multi-agent
AutoGen — Microsoft's multi-agent framework
Camel-AI — role-playing research framework
ChatDev — sister project with waterfall roles

Frequently asked questions

What's MetaGPT best at?

Greenfield, well-scoped apps — games, CRUD sites, scripts — where the SOP produces a coherent artefact. It struggles on open-ended requirements and large existing codebases.

MetaGPT vs CrewAI?

MetaGPT ships opinionated, hard-coded SOPs for software roles. CrewAI is a generic multi-agent library where you design the roles yourself. MetaGPT wins on 'spec-to-app'; CrewAI wins on flexibility.

Sources

MetaGPT — docs — accessed 2026-04-20
MetaGPT GitHub — accessed 2026-04-20