MetaGPT

An agent is a Role that owns a list of Actions and loops through observe, think and act. Roles do not call each other directly: each one declares which message types it reacts to with _watch, and publishes its output as a Message into a shared environment that other roles observe.

A Team wraps that environment. You hire roles, set a dollar budget with invest, submit the requirement with run_project, and run(n_round=...) rounds. The bundled software-company SOP passes work from product manager to architect to project manager to engineer, with optional code review and QA.

Each role's memory is the list of messages it has observed (rc.memory), read back with get_memories(k). Generated artifacts are written to a project workspace, and the CLI can resume a project from serialized storage or work incrementally on an existing repository.

Setting is_human=True on a role (for example a reviewer) hands that role to a person: when its turn comes, the run pauses for terminal input and the reply is passed to the other agents as that role's message. The docs state two limits: input is terminal-only, and the human must follow the same content format the downstream logic expects from an LLM. Custom _act code must call actions registered via set_actions for this to work. No approval gate for individual tool calls is documented.

pip install --upgrade metagpt

import asyncio
from metagpt.roles import Architect, Engineer, ProductManager, ProjectManager
from metagpt.team import Team

async def main(idea: str):
    company = Team()
    company.hire([ProductManager(), Architect(), ProjectManager(), Engineer()])
    company.invest(investment=1.0)   # budget in US dollars for model calls
    company.run_project(idea=idea)
    await company.run(n_round=4)

# Needs ~/.metagpt/config2.yaml (create it with: metagpt --init-config)
asyncio.run(main("A command-line tool that converts CSV files to Markdown tables"))

• The README requires Python 3.9 or later but below 3.12.
• Run metagpt --init-config and put the model key in ~/.metagpt/config2.yaml before the first run.
• The README asks for Node.js and pnpm before actual use; diagram output needs a Mermaid engine such as @mermaid-js/mermaid-cli.
• The docs estimate about $0.2 of GPT-4 fees for an analysis-and-design example and about $2.0 for a full project (project's own estimate).
• Besides the PyPI package, the README lists installing from the Git repository (pip install --upgrade git+https://github.com/geekan/MetaGPT.git) or from an editable clone (pip install --upgrade -e .), and the installation docs describe a Docker image (metagpt/metagpt:latest).
• A closed issue (#1910) reports pip install metagpt==0.8.2 failing dependency resolution through a newer agentops; the reporter's workaround was pinning agentops==0.3.17.
• GitHub marks v0.8.1 as the latest release although v0.8.2 was published in March 2025.

Install check by this site (2026-09-30, uv venv --python 3.11 && uv pip install --upgrade metagpt, macOS arm64) failed: dependency resolution fails. metagpt 0.8.2 (latest on PyPI, requires Python >=3.9,<3.12) pins lancedb==0.4.0, and no 0.4.x release of lancedb is available on pypi.org as of 2026-09-30, so the resolver cannot select 0.8.2; uv then backtracked to metagpt 0.1, whose pandas 1.4.1 source build failed. Tested with Python 3.11.

The repository is not archived, but activity is low: the last push was on 2026-01-21 and the newest tagged release is v0.8.2 from March 2025.

No official MCP, A2A or AG-UI support was found in the README, docs or code; all three are recorded as unknown.

Yes. The framework is MIT-licensed with no paid tier of its own; you pay your model provider. The same team sells MGX (mgx.dev now redirects to Atoms at atoms.dev), which its site presents as an AI website and app builder with free and paid plans rather than as a hosted MetaGPT service.

Python 3.9 or later but below 3.12, according to the README.

Each role watches for specific upstream message types and publishes its output to a shared environment, following an SOP; a Team sets the budget and number of rounds, and a human can replace any role via is_human=True.