MetaGPT

MetaGPT's core idea — encode a software company's standard operating procedure (Code = SOP(Team)) as a fixed pipeline of LLM roles (Product Manager → Architect → Project Manager → Engineer → QA), each with structured document handoffs instead of free-form chat — is a genuinely distinctive answer to the 'multi-agent systems degenerate into a chat loop' problem, and it's the one open GitHub issue from a would-be contributor (issue #2141) that a stranger volunteered unprompted in August 2026. The generate_repo() one-line-to-repo call and the separate Data Interpreter role (notebook/browser/shell tool use for data analysis and ML tasks) are both usable as installed today. The project's research output is real and independently checkable: the core MetaGPT paper was accepted as an oral at ICLR 2024, and follow-on work (AFlow, at ICLR 2025) also went to oral presentation — both listed on OpenReview under FoundationAgents/MetaGPT contributors.

The repository's last merge to main was 21 January 2026 (PR #1897, a Windows-terminal compatibility patch), and the last tagged PyPI release is 0.8.2, published 9 March 2025 — 18 months old as of this review. The GitHub repo still shows heavy inbound traffic (70,050 stars, 8,907 forks, 132 open issues), but recent issues and PRs are dominated by unrelated vendor pitches and speculative feature requests (e.g. a phone-pairing PR for an unrelated third-party product, an LLM-gateway vendor's unsolicited config recipe) with no maintainer replies visible in the thread history. This pattern is explained by the vendor's own account of its roadmap: MetaGPT's official successor page (atoms.dev/metagpt) describes the team's trajectory explicitly as 'MetaGPT (2023, open-source framework) → MGX (2024, commercial product) → Atoms (now, complete business platform)' — the company that built MetaGPT has moved its engineering effort to a paid, closed product and treats the open-source repo as a legacy artifact.

Three GitHub Security Advisories are open against MetaGPT ≤0.8.2, all reported through public issues in 2026 with no maintainer response noted in either the advisory or the linked issue thread: GHSA-xr7v-m9px-q4qj / CVE-2026-6110 (CVSS 7.3) — an eval() call on unvalidated LLM output in the Tree-of-Thought solver (metagpt/strategy/tot.py) that a prompt-injected or malicious model response can turn into arbitrary code execution; GHSA-r5v8-c28h-f8r8 / CVE-2026-6111 — server-side request forgery via decode_image(); and GHSA-3c3g-7hwg-9qmr / CVE-2026-10566 (CVSS 5.3) — a deserialization flaw in Message.check_instruct_content(). A fix for the eval() issue was submitted as PR #1946 in February 2026 but sits unmerged and open six months later. A related eval() pattern in a different file (serialize.py) was patched previously, per the reporting issue, so the project has fixed this class of bug before — just not this instance.

MIT license, confirmed on the GitHub repo and the PyPI package metadata — no seat fees, no usage tier; you run it against your own LLM API keys (OpenAI, Anthropic, Azure, and roughly a dozen others are supported via config2.yaml). There is no hosted/paid version of MetaGPT itself; the commercial descendant is a separate product (Atoms, formerly MGX) with its own pricing, not documented here.