Authority AI Editorial
MetaGPT: an agent team that works like a software company
MetaGPT is an open-source multi-agent framework that takes one line of requirements and runs it through a simulated software company: a product manager agent writes user stories, an architect agent designs the system, a project manager agent breaks down tasks, and engineer agents write the code. It is for developers who want more structure than a single coding agent replying to a prompt, and who are willing to set up a Python project to get a full pass of documents and code out of one request. The project was originally published as geekan/MetaGPT and now lives under the FoundationAgents organization on GitHub; this article uses the current name.
What it does
Feed MetaGPT a one-sentence requirement, such as building a small game or a utility script, and it runs that requirement through a sequence of role-based agents rather than a single model call. Each agent has a defined job and produces a defined artifact: user stories and competitive analysis from the product manager role, a system design and API spec from the architect role, a task breakdown from the project manager role, and working code from the engineer role. The framework's own description of this is "Code = SOP(Team)," meaning it tries to encode the standard operating procedures a real software team would follow and hand them to a team of language-model agents instead of a team of people. Beyond the software-company pipeline, MetaGPT also includes a Data Interpreter component aimed at code-writing and data-analysis tasks that do not need the full multi-role process.
How you use it
MetaGPT is a Python package, and the current README specifies Python 3.9 or later but earlier than 3.12. The straightforward install is pip install --upgrade metagpt, or you can clone the repository and run an editable install with pip install --upgrade -e . if you want to modify the source. After installing, metagpt --init-config sets up your configuration, mainly your model API keys, and then you can run it directly from the command line, for example metagpt "Create a 2048 game", or import it as a library inside your own Python code if you want to call the multi-agent pipeline programmatically rather than from a terminal.
What to know before adopting it
MetaGPT is MIT licensed. Its approach is documented in a peer-reviewed paper, "MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework," presented at ICLR 2024, so the multi-role design is not just marketing copy, it is a research result you can read and cite. A related paper from the same group, "AFlow: Automating Agentic Workflow Generation," was accepted as an oral presentation at ICLR 2025. Because MetaGPT runs several agent roles in sequence for a single request, expect it to make more model calls, and take longer, than a single-agent coding tool for the same task, and expect the output to look like project documents plus code rather than just code. There is no separate hosting or subscription cost from the project itself; you pay for whatever model API calls the pipeline makes with the credentials you configure. Note the org move from geekan to FoundationAgents if you are looking at older tutorials or blog posts that still reference the previous path.
Where it sits next to other agent tools
MetaGPT sits closer to project-generation frameworks than to a general coding assistant. Where a tool like OpenHands or Claude Code is built to work inside an existing repository on an ongoing basis, MetaGPT is built to take a request from zero to a structured first pass, complete with the planning documents a team would normally produce before writing code. It overlaps with orchestration frameworks like AutoGen in that both run multiple agents in a conversation, but MetaGPT's roles and outputs are fixed around the software-company metaphor rather than being a general-purpose conversation pattern you configure yourself. If you want a quick prototype plus the supporting documents, MetaGPT is a reasonable starting point; if you are iterating inside a codebase you already have, a single coding agent is usually the more direct tool. For either kind of agent, ready-made instructions can save setup time: the AI Ultimate Skill Collection packages 3,029 ready-made skills in 12 packs, written procedures an agent follows step by step, whether you are running MetaGPT's role-based pipeline or a single agent against a live repository.