MetaGPT

Short answer. MetaGPT is an open-source multi-agent framework by DeepWisdom (now FoundationAgents) that assigns the roles of a software company — product managers, architects, project managers and engineers — to LLMs, so a one-line requirement produces user stories, requirements, data structures, APIs and code under the philosophy “Code = SOP(Team)”.

Key facts.

• Role-based multi-agent architecture: a simulated software company with PM, architect, project-manager and engineer agents, orchestrated by standard operating procedures.
• Consumed as a CLI (metagpt "Create a 2048 game") or a Python library (from metagpt.software_company import generate_repo), with a Data Interpreter for code-and-data tasks.
• Model-agnostic: configurable backend (OpenAI, Azure, Ollama, Groq and other API-compatible providers), configured in ~/.metagpt/config2.yaml.
• MIT-licensed and free; the user supplies their own model API keys.
• Targets Python 3.9–3.11 (3.12 not yet supported per the README).
• The productized natural-language-programming layer is MGX (MetaGPT X) at mgx.dev, described as “the world’s first AI agent development team.”
• Repository moved from geekan/MetaGPT to FoundationAgents/MetaGPT.

What this means. MetaGPT is the reference implementation of the “multi-agent company” idea — a framework that encodes a software team’s SOPs into a multi-agent pipeline, whose commercial energy has shifted to a productized layer (MGX) rather than the SDK itself.

What is uncertain. The reason for the repo move, the company’s headquarters and funding, Python 3.12 timing, the framework’s production reliability, and the formal relationship between the open framework and MGX are not documented.

Sources.

evidence_id	source_name	source_url	source_type	published	verified	confidence	conflict
src-agents-metagpt-1	MetaGPT GitHub repository (FoundationAgents)	https://github.com/FoundationAgents/MetaGPT	Official documentation	—	2026-09-30	high	—
src-agents-metagpt-2	MetaGPT documentation (DeepWisdom)	https://docs.deepwisdom.ai/	Official documentation	—	2026-09-30	high	—
src-agents-metagpt-3	MGX (MetaGPT X) product site	https://mgx.dev/	Official	—	2026-09-30	high	—
src-agents-metagpt-4	MetaGPT project page (Foundation Agents)	https://foundationagents.deepwisdom.ai/projects/metagpt/	Official	—	2026-09-30	high	—

Why it matters

MetaGPT matters as the framework-level contrast to the platform agents in this database. Where Coze and Dify are managed or self-hosted platforms, MetaGPT is a library — a research-grade multi-agent SDK that a developer composes into their own system.

China AI Hub analysis indicates MetaGPT’s structural role is idea-source rather than deployment target: its “software company” SOP model and its research lineage (the SPO, AOT and AFlow papers) have shaped how the broader agent field thinks about role decomposition and self-optimizing workflows, even as its direct production use remains a developer’s assembly job rather than an out-of-the-box product. That is why its commercial energy has moved to MGX, which productizes the same multi-agent idea behind a natural-language interface.

China AI Hub analysis: this split — an MIT framework that is free and self-hosted, plus a commercial product that wraps it — is the same open-core pattern that Dify and FastGPT run, but applied to a framework rather than a platform. The difference is material: Dify and FastGPT are tools you run; MetaGPT is a library you import. Its influence therefore shows up less in deployments than in the design patterns (role-based SOPs, self-optimizing workflows) that the rest of the agent field has adopted.

How it differs from the platforms and frameworks

MetaGPT sits in a distinct cell of the agent ecosystem.

• Dify and FastGPT are platforms — self-hostable application builders (workflow/RAG/observability, or knowledge-base/RAG) that an operator runs to build applications.
• DeepSeek Harness is a plugin-composed runtime whose agent loop is replaceable; it defaults toward a vendor’s models.
• MetaGPT is a role-based multi-agent library — a pip-installable SDK whose core claim is not “run this to build apps” but “encode your team’s SOPs as agents and compose them in Python.”

China AI Hub analysis: the cleanest contrast is framework-versus-platform. A platform gives you a running system and a UI; a framework gives you building blocks and a loop. MetaGPT is the clearest example in this database of the latter — a developer-facing library whose most famous artifact is an idea (the software-company SOP) rather than a hosted surface, which is why it pairs naturally with a productized layer (MGX) rather than replacing one.

Practical implications

For researchers and framework builders. MetaGPT is the reference to study role-based SOP orchestration — the SPO/AOT/AFlow lineage is a research asset in its own right, and the MIT license makes the code freely inspectable.

For developers wanting multi-agent codegen. The CLI and library paths are documented, but the developer must supply and pay for their own LLM keys and stay on Python 3.9–3.11 — a research-grade setup, not a turnkey product.

For teams wanting a product. MGX is the productized path; the framework itself is the SDK underneath. Anyone evaluating MetaGPT for production should be clear which layer they are actually adopting.

What the evidence shows

The evidence is strong on the framework’s conceptual design because it is explicit in primary material: “Assign different roles to GPTs to form a collaborative software entity for complex tasks,” with the PM/architect/project-manager/engineer roles and the “Code = SOP(Team)” philosophy documented across the repo and docs. The model-agnostic backend (OpenAI, Azure, Ollama, Groq) and the config2.yaml mechanism are documented, and the MGX product is described as the natural-language-programming layer built on the framework.

The gaps are provenance and evaluation. The repo move from geekan/MetaGPT to FoundationAgents/MetaGPT is recorded but not explained; the company’s headquarters and funding are not public; Python 3.12 timing is unstated; and the framework’s production reliability is not independently benchmarked. China AI Hub analysis indicates MetaGPT is a case where conceptual influence clearly outruns measurable deployment — its ideas are widely cited in the agent field, but no independent production-evaluation evidence is recorded in this database.

Where this fits

Workload	Relevance
Multi-agent software development from one requirement	High
Role-based SOP orchestration (PM/architect/engineer)	High
Research prototyping (SPO, AOT, AFlow)	High
Data analysis via Data Interpreter	High
Natural-language programming via MGX	Moderate (separate product)
Turnkey hosted deployment	Low (framework, not a product)
Production multi-tenant SaaS on the framework	Low (research-grade SDK)

Relevance judgments are China AI Hub analysis based on documented capabilities, not vendor claims.

Field reference

Field	Value	Evidence type
Underlying model	Model-agnostic (OpenAI, Azure, Ollama, Groq, API-compatible)	Official
Target users	Developers, researchers	Official
Platform	Python framework (pip install metagpt) + CLI	Official
OS	Cross-platform (Python 3.9–3.11)	Official
Browser / computer use	Not publicly documented	Not publicly documented
Coding	Yes (software-company code generation)	Vendor-reported
Autonomous task execution	Yes (multi-agent SOP orchestration)	Vendor-reported
MCP	Not publicly documented	Not publicly documented
Tool calling	Yes (Data Interpreter, tools)	Official
Memory	Yes (agent memory / workspace)	Official
Workflow	Role-based SOP: PM / architect / PM / engineer pipeline	Official
API	Yes (Python library API)	Official
Pricing	Free and open source (MIT); MGX offered separately	Vendor-reported
Region	Global (self-hosted)	Official
Open-source	Yes — MIT	Official
Deployment	Self-hosted	Official
Limitations	Python 3.9–3.11; user supplies model keys; research-grade SDK	Official
Source	GitHub · docs	Official
Last verified	2026-09-30	Official

See the DeepWisdom company profile, the choosing-an-agent guide, and the site’s AI agents and tool calling technology pages. For the structural reading of MetaGPT as framework-versus-platform, see the research on the agent ecosystem structure and the rise of Chinese AI agents.

Labels used above: Official fact (from the MetaGPT GitHub repository, DeepWisdom documentation and the Foundation Agents project page), Vendor-reported claim (capability statements by DeepWisdom), and China AI Hub analysis (our synthesis, always introduced as such). No third-party evaluation evidence is currently recorded for the MetaGPT framework.