MetaGPT

Overview

MetaGPT is an open-source Python framework for building software-company style multi-agent workflows. Its README frames the core pattern as assigning different roles to GPT-backed agents so they collaborate on complex tasks. For software generation, MetaGPT can take a one-line requirement and produce artifacts such as user stories, competitive analysis, requirements, data structures, APIs, documents, and repository files.

The entry is useful for Claude-adjacent teams because MetaGPT is a high-search, high-star agent framework with a distinct mental model: instead of a single coding assistant or generic tool-calling loop, it models a team with product manager, architect, project manager, engineer, and related roles following orchestrated SOPs.

Install

Install the PyPI package:

pip install --upgrade metagpt

The README also documents conda setup, source installs, and a requirement to install Node.js and pnpm before actual use:

conda create -n metagpt python=3.9
conda activate metagpt

Initialize configuration before running workflows:

metagpt --init-config

Then edit ~/.metagpt/config2.yaml with the selected LLM provider, model, base URL, and API key.

Common Commands

Generate a small project from a requirement:

metagpt "Create a 2048 game"

Use MetaGPT as a Python library:

from metagpt.software_company import generate_repo

repo = generate_repo("Create a 2048 game")
print(repo)

Run a Data Interpreter workflow:

import asyncio
from metagpt.roles.di.data_interpreter import DataInterpreter


async def main():
    di = DataInterpreter()
    await di.run("Run data analysis on sklearn Iris dataset, include a plot")


asyncio.run(main())

Agent Capabilities

MCP Fit

MetaGPT is not an MCP server. It belongs in tools as a framework and CLI for role-based multi-agent workflows. It can still be useful in the same search cluster as MCP because teams often compare MCP tool access with higher-level agent frameworks. If MetaGPT agents are connected to external tools, browsers, repos, email, or RAG systems, apply the same permission and approval discipline you would apply to MCP-connected agents.

Source Review

Verified on 2026-06-18:

• GitHub reports FoundationAgents/MetaGPT as an MIT-licensed repository with 68,000+ stars and a latest release tag v0.8.1.
• The README describes MetaGPT as a multi-agent framework that assigns different roles to GPTs and models a software company as a multi-agent system.
• The README documents pip install --upgrade metagpt, Python 3.9+ and <3.12, Node.js and pnpm prerequisites, metagpt --init-config, ~/.metagpt/config2.yaml, CLI usage, Python library usage, and Data Interpreter usage.
• setup.py declares the metagpt package, MIT license, Python >=3.9,<3.12, console script metagpt=metagpt.software_company:app, and extras for Selenium, search, RAG, tests, development, and Android assistant workflows.
• requirements.txt includes provider, browser, RAG/vector, GitHub, email, notebook, data, and model integration dependencies.
• PyPI resolves package metadata for metagpt version 0.8.2; current main source metadata shows a newer setup.py version value, so version-sensitive install decisions should check PyPI and the upstream repository before pinning.

Safety and Privacy

MetaGPT's most attractive workflow is also the risky one: a brief requirement can produce many files, diagrams, dependencies, and design decisions. Keep that output in a reviewable workspace, run tests before trusting it, and do not let generated repositories inherit production credentials or privileged shells.

Role messages, product requirements, architecture documents, generated source, notebook outputs, and tool results may expose sensitive product plans or customer context. Treat ~/.metagpt/config2.yaml, generated workspaces, logs, and shared artifacts as private until reviewed.

Duplicate Check

Checked current content/tools/, content/mcp/, content/agents/, content/skills/, guides, collections, open pull requests, and repository-wide content for FoundationAgents/MetaGPT, geekan/MetaGPT, MetaGPT, metagpt, AI software company, software company as multi-agent system, Data Interpreter, natural-language programming, and role-based multi-agent framework. Existing entries cover adjacent agent frameworks and coding agents such as AutoGen, CrewAI, LangGraph, Agno, Mastra, Pydantic AI, Smolagents, CAMEL, OpenHands, Open SWE, mini-SWE-agent, and Claude Code, but no dedicated MetaGPT tools entry, exact source URL duplicate, target file, or open duplicate PR was found.