Purpose
MetaGPT is a multi-agent framework (The Multi-Agent Framework), described as "the First AI Software Company," with the goal of advancing toward natural language programming. It assigns different roles to different GPTs, forming a collaborative entity to handle complex tasks; the input is a one-line requirement, and the output includes user stories, competitive analysis, requirements, data structures, APIs, documents, and more.
Main Capabilities
- Organizes the process in the manner of "a software company as a multi-agent system": internally it includes roles such as product manager, architect, project manager, and engineer, provides the complete process of a software company, and is equipped with a carefully orchestrated SOP.
- The core idea is Code = SOP(Team): materialize the SOP and apply it to a team composed of LLMs.
- Takes a one-line requirement as input and produces deliverables such as user stories, competitive analysis, requirements, data structures, APIs, and documents.
- The available use cases/examples listed in the repository include: Data Interpreter, Debate, Researcher, Receipt Assistant.
- Data Interpreter can be used to write code.
Usage
Environment requirements: ensure the system has Python 3.9 or higher installed, but lower than 3.12; you can use python --version to check. The official example uses conda to create the environment:
conda create -n metagpt python=3.9 && conda activate metagpt
Before actual use, node and pnpm need to be installed. For detailed installation instructions, refer to cli_install or docker_install.
You can initialize the MetaGPT configuration with the following command, or manually create the ~/.metagpt/config2.yaml file. You can configure ~/.metagpt/config2.yamlaccording to the examples and documentation.
After installation, you can use MetaGPT in the CLI, or use it as a library.
Applicable Scenarios
- Starting from natural language requirements, automatically generate documents and design artifacts related to software projects (user stories, competitive analysis, requirements, data structures, APIs, documents, etc.).
- Scenarios that require multiple AI roles to collaborate to complete complex tasks.
- Wanting to write code through Data Interpreter.
- Use cases officially listed and available for review: Data Interpreter, Debate, Researcher, Receipt Assistant.
Related Products and Updates
- March 10, 2025: mgx.dev became @ProductHunt's #1 Product of the Week.
- March 4, 2025: mgx.dev became @ProductHunt's #1 Product of the Day.
- February 19, 2025: Officially released the natural language programming product MGX (MetaGPT X) — described as the world's first AI agent development team.
- February 17, 2025: Introduced two papers, SPO and AOT; the code is available for review.
- January 22, 2025: The paper AFlow: Automating Agentic Workflow Generation was accepted by ICLR 2025 as an oral presentation (top 1.8%), ranking 2nd in the LLM-based Agent category.
Resources and Limitations
- Officially provides online documentation, QuickStart, and demo videos, and it can be tried on the MetaGPT Huggingface Space.
- There is a development roadmap, contribution guidelines, and a Discord channel; contributors can fill out a form.
- Contact: email alexanderwu@deepwisdom.ai; for technical questions, you can also open a new issue in the GitHub repository. The official statement says all questions will be answered within 2-3 working days; you can follow @MetaGPT_ on Twitter for the latest R&D updates; paper citations are provided in BibTeX.
- In terms of limitations: the materialsonly give prerequisites such as the Python version and node/pnpm, and do not provide more information such as fees, prices, platforms, or results; some content (such as specific CLI commands, specific configuration items in config2.yaml, and detailed steps for each use case) is not expanded in the provided materials and is not yet available.
