Automate literature review and code verification with a multi-agent CLI
Imagine throwing a new research topic at a system where multiple agents simultaneously find literature, verify papers, and write drafts. This tool is a command-line program that automates complex research processes. It generates files with structured research results and verifiable sources, so you do not have to write code manually or read papers by hand.
Automating the research workflow
The tool provides several commands that assist with the entire research process, going beyond simple search. /deepresearch deploys researchers in parallel to investigate topics and check citations, while /lit summarizes consensus, disagreements, and unresolved issues through literature reviews. /review offers critiques and revision plans for papers or drafts, and /audit compares paper claims with public code. Additionally, /replicate creates replication plans, and /recipe displays machine learning training recipes, including datasets and code, ranked by order.
Data sources and agent structure
The tool searches various academic databases such as alphaXiv, Semantic Scholar, arXiv, PubMed, and Hugging Face, as well as the web and local PDFs. Internally, four agents—researcher, verifier, reviewer, and writer—collaborate to perform tasks. Setting up a free API key allows you to avoid shared rate limits, and the tool can integrate with local models or editors like Zed. All results are saved in the outputs/ folder, and some commands generate source files containing the sources used and verification details.
Installation and initial setup
You can install the tool using the curl command on macOS and Linux, or the PowerShell command on Windows. Global installation via npm is also possible in environments with Node.js 22.22 or higher. After installation, run the feynman setup command to log in to a model provider, and the tool is ready to use. The installation guide includes information on pinned version management, removal methods, and skill-only bundles for Codex and OpenCode.
Data collection and contribution
The tool sends anonymous usage data to PostHog. The transmitted content includes commands, workflows, tools, models, token counts, error messages, and stack traces, while prompts, model outputs, paper content, and tool arguments are never sent. You can disable this feature by setting the FEYNMAN_TELEMETRY=off environment variable. To contribute, clone the repository, install dependencies with npm, and run the tests. This project is released under the MIT license.