🏢 MetaGPT：67k Stars 的多智能体框架，一句需求自动生成整个软件公司

老实说，你有没有想过这么一个场景：你泡杯咖啡的功夫，AI 就把产品需求文档、架构设计、API 文档、甚至完整代码都给你整出来了？这不是画饼，MetaGPT 干的事就是——把整家软件公司的角色塞进 GPT，一句话需求扔进去，出来的是一个完整项目。

项目地址：github.com/FoundationAgents/MetaGPT | ⭐ 67,967 | 🛠 Python | 作者：geekan/MetaGPT → FoundationAgents

不是"一个"Agent，是一整个团队

市面上大多数 Agent 框架解决的是"一个 AI 怎么干活"，MetaGPT 的思路完全不同——它模拟了一整家软件公司：

最骚的操作是这些角色不是各干各的，它们之间有完整的 SOP（标准作业流程）串联。Code = SOP(Team) 是它的核心哲学——把人类软件开发的管理流程，原封不动地搬到 LLM 协作中。

一行命令，一个项目

安装很简单，Python 3.9+ 直接 pip：

pip install --upgrade metagpt
# 还需要安装 node 和 pnpm（用于生成前端代码）

配好 LLM API 密钥：

# ~/.metagpt/config2.yaml
llm:
  api_type: "openai"
  model: "gpt-4-turbo"
  base_url: "https://api.openai.com/v1"
  api_key: "YOUR_API_KEY"

然后一句话启动：

metagpt "Create a 2048 game"

它会自动在 ./workspace 下生成一个完整项目。你也可以在 Python 里调用：

from metagpt.software_company import generate_repo
from metagpt.utils.project_repo import ProjectRepo

repo: ProjectRepo = generate_repo("Create a 2048 game")
print(repo)  # 打印项目结构

不只是生成代码

MetaGPT 还有个叫 Data Interpreter 的模块，专门做数据分析和可视化：

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())

学术硬实力

MetaGPT 不只是个开源玩具——它的论文 AFlow: Automating Agentic Workflow Generation 被 ICLR 2025 接收为 oral presentation（top 1.8%），在 LLM-based Agent 类别排名第 2。2025 年 3 月还推出了 mgx.dev，全球首个自然语言编程产品。

要点总结

🏢 MetaGPT: 67k Stars Multi-Agent Framework — Turn One Requirement Into a Full Software Company

Honestly, imagine this: you sip your coffee while AI generates product requirement docs, system architecture, API specs, and even full production code. MetaGPT takes a single-line requirement and outputs a complete project — because it doesn't just run one agent, it simulates an entire software company inside GPT.

Repo: github.com/FoundationAgents/MetaGPT | ⭐ 67,967 | 🛠 Python | License: MIT

Not One Agent — A Whole Team

While most agent frameworks solve "how one AI works," MetaGPT simulates real company roles:

These roles don't work in isolation — they follow SOPs (Standard Operating Procedures) just like a real dev team. The core philosophy: Code = SOP(Team).

One Command, Full Project

pip install --upgrade metagpt

Configure your LLM API:

# ~/.metagpt/config2.yaml
llm:
  api_type: "openai"
  model: "gpt-4-turbo"
  base_url: "https://api.openai.com/v1"
  api_key: "YOUR_API_KEY"

Run it:

metagpt "Create a 2048 game"

Or use as a Python library:

from metagpt.software_company import generate_repo
repo: ProjectRepo = generate_repo("Create a 2048 game")
print(repo)

Beyond Code Generation

MetaGPT's Data Interpreter module handles data analysis and visualization:

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())

Academic Credentials

MetaGPT's paper "AFlow: Automating Agentic Workflow Generation" was accepted as an oral presentation (top 1.8%) at ICLR 2025, ranking #2 in the LLM-based Agent category. In March 2025, they launched mgx.dev — the world's first natural language programming product.

Key Takeaways