Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

MetaGPT

Python framework that runs a one-line request through LLM roles modelled on a software company, now largely in maintenance mode.

Reviewed by ReviewRamble editors · Updated

View on GitHub Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.2/5

MetaGPT is the paper that made role-playing agent teams famous and a library that has barely moved since, so read it for ideas and build on something maintained.

What is MetaGPT?

MetaGPT turns a one-line requirement into the paperwork and code a small software company would produce. LLM agents play product manager, architect, project manager and engineer, each following a written procedure (the project's motto is Code = SOP(Team)), and hand documents to one another: user stories, competitive analysis, a requirements list, data structures, API designs and finally a repository in ./workspace. `metagpt "Create a 2048 game"` is the README's example.

The same library ships Data Interpreter, an agent that plans a data-analysis task, writes Python and runs it in a local Jupyter kernel, plus research code from the team behind it, such as AFlow (automated workflow generation, an ICLR 2025 oral) and SPO prompt optimisation. Models are configured in ~/.metagpt/config2.yaml: OpenAI, Azure, Ollama, Groq, Bedrock and others.

The company behind it, DeepWisdom, now puts its effort into Atoms, a closed-source hosted app builder that launched as MGX (MetaGPT X) in February 2025 and was renamed in January 2026. The open-source framework is MIT licensed and still installs from PyPI, but its last release, v0.8.2, dates from March 2025.

Is MetaGPT worth it?

MetaGPT matters more as a reference design than as a dependency. Its ICLR 2024 paper showed that giving agents fixed roles and structured hand-offs (a PRD, then a design, then tasks) scored better in its benchmarks than freer agent chat, and that idea now runs through CrewAI, ChatDev-style demos and most spec-driven coding tools. The repository still has about 71,000 GitHub stars.

The activity tells a different story. The last tagged release is v0.8.2 from 9 March 2025. In the 12 months to November 2026 the main branch took one merged change, Windows Terminal support on 21 January 2026, while 152 pull requests sit open. The package still pins Python to 3.9 through 3.11, so it will not install on 3.12 or 3.13 without workarounds, and the sample config still points at gpt-4-turbo.

The reason is commercial. DeepWisdom launched MGX in February 2025 and renamed it Atoms on 13 January 2026; mgx.dev now redirects to atoms.dev. KrASIA reports the company says MGX reached about 500,000 registered users in its first month and that it has raised about $30.8 million. Atoms is a hosted, closed product with credit plans from $20 a month, and the open repository shows little sign of the agent work that went into it.

The catch: the software-company pipeline is built for new projects. It starts from a one-line requirement, and each role works from the previous role's documents rather than from an existing codebase, so it has little to offer on a mature repository. Data Interpreter executes the code it writes in a local Jupyter kernel with your user's permissions, so run it in a container. Token use is high: one request passes through four or more roles, each a full model call with the accumulated documents.

Our take: MetaGPT is a good afternoon's reading for anyone designing a multi-agent system, and its examples folder is a useful catalogue of research ideas, but for production work the maintained frameworks have taken its central idea and kept shipping.

Where it earns its place

  • A PhD student working on LLM agent coordination

    Reproduce a well-known multi-agent baseline for a paper

    Install metagpt 0.8.2 in a Python 3.11 environment, run the software-company pipeline on benchmark prompts and compare against the ICLR paper. The code is free; the cost is API tokens across four or more role calls per prompt.

  • A tech lead at a 20-person consultancy

    Turn a client's one-paragraph brief into a first PRD and system design for a pitch

    Run `metagpt "<brief>"` and keep the generated user stories, competitive analysis and API design as a starting draft; discard the code. A person still edits every document before it reaches the client.

  • A data analyst at a mid-sized retailer

    Get a quick exploratory analysis and charts from a CSV

    Data Interpreter plans the steps, writes pandas and plotting code and runs it in a Jupyter kernel. Run it inside a container, because the generated code executes with your permissions.

Before you commit

Setup
pip install metagpt on Python 3.9 to 3.11, install Node.js and pnpm, then `metagpt --init-config` and add a model key to ~/.metagpt/config2.yaml; a Docker image is also documented.
Integrations
OpenAI, Azure OpenAI, Ollama, Groq, Bedrock and other providers through config2.yaml; usable as a CLI or as a Python library (generate_repo, DataInterpreter).
Data and privacy
Self-hosted library under MIT; prompts go only to the model provider you configure. Generated repositories land in ./workspace, and Data Interpreter runs code on the host. Atoms, the hosted product, is a separate service with its own terms.
Cost at scale
No licence cost; each request makes several full model calls with growing context, so token bills scale with the number of roles and review rounds rather than with users.
Limits
Last release March 2025, 152 open pull requests, no Python 3.12+ support, and a pipeline designed for small new projects rather than existing repositories.

What’s good

  • Clear, influential design: fixed roles passing structured documents
  • One command produces PRD, design, task list and code for a small app
  • Data Interpreter and research code (AFlow, SPO) in one package
  • MIT licence with no commercial limits

What’s not

  • No release since v0.8.2 in March 2025 and one merged change in a year
  • Python 3.9 to 3.11 only; 3.12 and 3.13 are unsupported
  • Built for new projects; roles work from documents, not your repository
  • Data Interpreter runs generated code locally without a sandbox
  • The team's new work goes into the closed Atoms product

Best for

Engineers and researchers studying how to structure multi-agent systems, not teams choosing a framework to run in production.

MetaGPT alternatives

MetaGPT: common questions

Is MetaGPT worth it?
MetaGPT is the paper that made role-playing agent teams famous and a library that has barely moved since, so read it for ideas and build on something maintained. We rate it 3.2 out of 5.
How much does MetaGPT cost?
Free and open source under the MIT licence; you pay your model provider. The team's hosted product Atoms (formerly MGX) is separate: Free with 15 credits a day, Pro from $20/month, Max from $100/month.
What are the downsides of MetaGPT?
No release since v0.8.2 in March 2025 and one merged change in a year; python 3.9 to 3.11 only; 3.12 and 3.13 are unsupported; built for new projects; roles work from documents, not your repository; data Interpreter runs generated code locally without a sandbox; the team's new work goes into the closed Atoms product.
Who is MetaGPT best for?
Engineers and researchers studying how to structure multi-agent systems, not teams choosing a framework to run in production.
How much does MetaGPT cost for a team or in production?
No licence cost; each request makes several full model calls with growing context, so token bills scale with the number of roles and review rounds rather than with users.
Is MetaGPT safe for company data?
Self-hosted library under MIT; prompts go only to the model provider you configure. Generated repositories land in ./workspace, and Data Interpreter runs code on the host. Atoms, the hosted product, is a separate service with its own terms.
What are the best MetaGPT alternatives?
The closest alternatives we have reviewed are CrewAI, AutoGen, LangGraph and OpenHands.

Answers checked Nov 20, 2026.