← Back to all projects

MetaGPT

Runs LLM agents as a software team that starts a project from scratch, not a helper for a codebase you already have.

OfficialMIT
Stars
71.1k
Forks
9.1k
Open issues
134
Last commit
21 Jan 2026

What is MetaGPT?

Hand MetaGPT a one-line requirement and a team of LLM roles works through it in turn, leaving a requirements document, a system design and source files in a folder under `workspace/`. Which roles show up depends on where you installed from, and the gap is not cosmetic: the published 0.8.2 package hires a product manager, an architect, a project manager and an engineer, while current main — the README's clone-and-install route — hires a team leader, a product manager, an architect, an engineer and a data analyst, with the project manager commented out. Under the demo sits a general framework: you subclass Role and Action, declare which upstream actions each role watches, and a shared Environment routes the messages that put the team in order. A separate agent, the Data Interpreter, does data work alone, planning the steps, writing Python and retrying what fails. The catch is scope — MetaGPT's own FAQ says functions start being left unimplemented past roughly 500 lines of generated code.

What can you do with MetaGPT?

  • One sentence in, a project folder out — metagpt "Create a 2048 game" fills a folder under workspace/ with the requirements document, the system design and the source files, and generate_repo() does the same from Python. Turning the design diagrams into pictures is a separate step that hands the diagram text to mmdc, installed with npm install -g @mermaid-js/mermaid-cli; without it you get a logged warning and the diagram source rather than a failed run, and mermaid.engine can be switched to playwright, pyppeteer, ink or none.
  • Roles, actions, and who watches whom — You subclass Role and Action, then call _watch to name the upstream actions a role reacts to. Every role observes a shared Environment and publishes its result back into it, so the order of work falls out of those subscriptions rather than from a graph you draw yourself.
  • Cap the spend, then pick the run back up — Team.invest(3.0) — or --investment, which defaults to $3 — sets a token budget that stops the run with an error once the accumulated cost reaches it, and --n-round (default 5) separately bounds the rounds of message passing. If the run then fails or you interrupt it, the whole team is written to disk and --recover-path picks it up from there.
  • The Data Interpreter plans, writes and runs Python — Ask for an analysis with a plot and it breaks the work into a plan, writes code for each step and executes it in a live Jupyter kernel, so variables survive from one step into the next. A step that raises is rewritten and rerun up to three times with the error text put back in front of the model. None of the software-company roles need to be around it.
  • Registering your own tool takes three steps, not one — @register_tool() on a function or class under metagpt/tools/libs turns its Google-style docstring into the schema the model reads. That alone does nothing: DataInterpreter.tools defaults to empty and the recommender that picks between tools is only built when it is non-empty, so you also import the module and construct DataInterpreter(tools=["magic_function"]).
  • The research the team publishes ships in the repo — examples/aflow searches a space of workflows represented as code with a Monte Carlo tree search, generating the agent workflow itself instead of you writing it; examples/spo rewrites prompts with no ground-truth labels; examples/sela applies the same tree search to automated machine learning. Each runs on its own, with no software company involved.

Before you choose MetaGPT

  • PyPI still serves 0.8.2 from March 2025, so the newer roles exist only on main, which has taken three commits since July 2025 and where a reported hang that stalls the engineer on Linux was still open in August 2026.Reported in#2110
  • Incremental mode expects a project MetaGPT itself created, and the guide puts requirements about architecture files or how one function should behave out of scope, so it is not a route into an existing repository.

Star history

17 Aug to 30 Oct · +1.2k

69.9k71.1k

Frequently asked questions

Is MetaGPT free for commercial use?

MetaGPT is released under the MIT licence — OSI-approved open source, which permits commercial use.

How can MetaGPT be deployed?

MetaGPT is available as Self-hosted / Runs locally.