MetaGPT is an open-source framework that models a software company with AI agents. Agents take on roles such as product manager, architect, project manager and engineer, and they follow fixed standard operating procedures (SOPs). From a one-line requirement, this produces a requirements document, a system design, a task list and code. The accompanying research paper was presented at ICLR 2024. This article explains the concept, how to get started, the current state of development and what MetaGPT is still good for today.
Fact sheet (as of October 2026)
| Attribute | Details |
|---|---|
| Origin | DeepWisdom and researchers from the FoundationAgents community |
| License | MIT |
| GitHub | around 70,000 stars (as of October 2026) |
| Language | Python 3.9 to 3.11, plus Node.js and pnpm for some features |
| Latest PyPI version | 0.8.2 from March 2025 |
| Commercial product | Atoms (called MGX or MetaGPT X until January 2026), a platform that builds applications from ideas described in natural language |
The concept: Code = SOP(Team)
MetaGPT's core idea is that human teams succeed because they have fixed processes and roles. MetaGPT applies this to agents:
| Role | Output |
|---|---|
| Product Manager | Requirements document (PRD) with user stories and competitive analysis |
| Architect | System design with data structures, interfaces and sequence diagrams |
| Project Manager | Task list with dependencies |
| Engineer | Code for the individual files |
| QA Engineer | Tests and review |
The agents do not exchange free-form chat messages. Instead, they exchange structured documents through a shared message pool that each role subscribes to as needed. This reduces misunderstandings and hallucinations at handoffs. The idea has influenced many later multi-agent frameworks.
Getting started
pip install --upgrade metagpt
metagpt --init-config # creates ~/.metagpt/config2.yaml
In the configuration file, you enter your model provider and API key. Then you start a project from the command line:
metagpt "Create a 2048 game in the terminal"
Or from Python:
from metagpt.software_company import generate_repo
from metagpt.utils.project_repo import ProjectRepo
repo: ProjectRepo = generate_repo("Create a 2048 game in the terminal")
print(repo)
The result ends up in the workspace folder with documents, diagrams and code.
Other features
- Custom roles and actions: You can define your own roles with actions, such as a research agent.
- Data Interpreter: An agent for data analysis that writes and runs code.
- Research projects: Work from the same group includes AFlow (automated workflow generation, ICLR 2025) and SPO (prompt optimization).
Current state of development
Since early 2025, the core team has focused mainly on the commercial product Atoms (formerly MGX, renamed in January 2026). The open-source repository still receives contributions, but the latest version on PyPI dates from March 2025, and supported Python versions stop at 3.11. For new production projects this is a risk: new model APIs, tooling standards such as MCP and security fixes in dependencies arrive slowly or not at all.
Strengths
- Clear role model with structured handoffs
- Complete artifacts: requirements, design and diagrams in addition to code
- Well-documented research as a foundation, useful for teaching and your own experiments
- MIT license
Limitations
- Infrequent releases and outdated Python support
- Generated projects are prototypes: MetaGPT is not designed for existing, large codebases. Coding agents such as OpenHands, Claude Code or Devin are a better fit there.
- High token costs, because every role produces detailed documents
- Rigid workflows: For tasks outside software development, you have to adapt a lot
MetaGPT compared
| MetaGPT | CrewAI | ChatDev | |
|---|---|---|---|
| Focus | Software development based on SOPs | General role-based teams and flows | Software development as a virtual company |
| Handoffs | Structured documents | Tasks with an expected output | Role-based chats |
| Maintenance | Few releases since 2025 | Active, frequent releases | Research project |
For the comparison with AutoGen, see MetaGPT vs. AutoGen.
Who is MetaGPT for?
- Research and teaching on multi-agent collaboration
- Experiments and prototypes where you want to turn an idea into a small project quickly
- Inspiration for your own systems with roles and structured handoffs
For production multi-agent systems, we recommend actively maintained frameworks such as LangGraph, CrewAI or the Microsoft Agent Framework.
Frequently asked questions
Is MetaGPT still being developed?
Only to a limited extent. The repository still receives contributions, but the latest version on PyPI is 0.8.2 from March 2025. The team behind MetaGPT focuses on the commercial product Atoms (formerly MGX).
What is the difference between MetaGPT and Atoms?
MetaGPT is the open-source Python framework under the MIT license that you install yourself. Atoms is a hosted commercial platform that builds websites and applications from descriptions in natural language. Atoms was called MGX until January 2026.
Which Python version do I need for MetaGPT?
Python 3.9 to 3.11. The latest published version does not support newer Python versions.
Is MetaGPT suitable for existing codebases?
Hardly. MetaGPT creates new, small projects from a requirement. For changes to existing code, coding agents such as OpenHands or Claude Code are a better fit.