Feynman: An Open-Source AI Agent That Verifies Research Papers Against Their Code

Feynman is an open-source, terminal-based AI research agent that reads scientific papers, checks their claims against actual code repositories, and produces cited research briefs. The video from Github Signals walks through the GitHub README for advaitpaliwal/feynman, pitching it as a fix for information overload in scientific literature. The core claim holds up on inspection. The repo is real, MIT-licensed, has around 9.2k stars, and does what the video describes: multi-agent research with source verification rather than plain summarization.

What Feynman Does

The transcript describes Feynman as a tireless lab assistant. You ask it to review a topic, and it searches databases, compares claims against code repositories, and cites its sources. The README confirms this. Feynman answers natural-language questions like "what do we know about scaling laws" by pulling from papers and web sources, then returns a research brief with inline citations. The terminal screenshot shown throughout the video displays exactly this flow, with a query about the latest research on scaling laws expanding into multiple search queries across web and academic sources.

The Four-Agent Architecture

The video highlights that Feynman runs four distinct internal agents to research, review, write, and verify facts. The repo lists them by name: a researcher that gathers evidence, a reviewer that critiques, a writer that structures drafts, and a verifier that manages citations. The terminal screenshot in the video shows these four agents listed in the session panel, running on Claude (the screenshot shows an Anthropic model string in the config). Splitting the work this way is what separates the output from a single-pass summary, since claims get checked by a different agent than the one that wrote them.

Research Workflows Beyond Summarization

The README screenshot visible in the frames lists a set of slash-command workflows: /audit checks a paper's claims against its public codebase for mismatches and reproducibility risks, /autoresearch runs autonomous experiment loops, /compare builds agreement matrices across multiple sources, /deepresearch runs a thorough source-heavy investigation, /lit produces literature reviews, /replicate builds replication workflows for papers and benchmarks, and /watch sets up recurring monitoring on a topic. The /audit workflow is the one the video leads with, and it is the most distinctive: comparing what a paper says against what its code actually does.

Installation and Licensing

The frames show a one-line installer for macOS and Linux (curl -fsSL https://feynman.is/install | bash) and a PowerShell equivalent for Windows. The installer ships a standalone bundle with a pinned Node.js runtime and verifies the release SHA-256 before replacing an existing install. The project is MIT-licensed, so it is free to use and modify. It is built on the Pi agent runtime and alphaXiv paper search, and integrates more than 40 scientific databases including PubMed, bioRxiv, ChEMBL, and UniProt.

Who This Is For

The video targets developers and scientists, claiming the tool turns hours of manual reading into minutes of automated analysis. That framing fits the feature set. Anyone doing literature review, checking whether an ML paper's results match its released code, or tracking a fast-moving research area gets the most value. The citation-first design matters here, since a brief you can trace back to sources is auditable in a way a chat summary is not.

Key Takeaways

Feynman is a real, MIT-licensed open-source AI research agent at advaitpaliwal/feynman with roughly 9.2k stars.
It uses four internal agents (researcher, reviewer, writer, verifier) so claims get checked separately from where they are written.
The /audit workflow compares a paper's claims against its public codebase to flag mismatches and reproducibility risks.
It integrates 40+ scientific databases including PubMed, bioRxiv, ChEMBL, and UniProt, and cites every claim in its briefs.
Installation is a single shell command on macOS, Linux, or Windows, with SHA-256 verification of releases.