Feynman

Feynman is an open source AI research agent built by Companion and distributed under the MIT License. It runs entirely on your local machine as a CLI tool, orchestrating multiple specialized agents to search papers, synthesize findings, run experiments, and produce fully cited research outputs. The project is hosted at github.com/companion-inc/feynman and as of May 2026 has reached version 0.2.58.

What It Is

Feynman is a terminal-based research agent that accepts natural language questions or structured slash commands and returns source-grounded answers. It is designed for AI/ML researchers and practitioners who need to move quickly from a research question to a cited brief, literature review, paper audit, or replication experiment. The tool is built on the Pi agent runtime and integrates with AlphaXiv for paper search and analysis, Hugging Face Hub for dataset and model inspection, and cloud GPU providers for compute-heavy tasks.

How the Agent Team Works

Feynman dispatches four bundled agents automatically based on the task:

• Researcher — gathers evidence across papers, the web, repositories, and documentation
• Reviewer — performs simulated peer review with severity-graded feedback and a revision plan
• Writer — structures research notes into briefs, drafts, and paper-style output
• Verifier — checks every citation, verifies source URLs, and removes dead links

The multi-agent /deepresearch workflow runs parallel researchers, synthesizes findings, and passes results through the Verifier before producing output.

Workflows and Slash Commands

Users can invoke workflows with natural language or explicit slash commands:

• /deepresearch — source-heavy multi-agent investigation
• /lit — literature review with consensus mapping and open questions
• /review — simulated peer review with severity scores
• /audit — paper-to-code mismatch audit for reproducibility claims
• /replicate — replication plan and sandboxed Docker execution
• /recipe — ranked ML training recipes grounded in papers, datasets, docs, and code
• /compare — side-by-side source comparison matrix
• /draft — polished paper-style draft with inline citations
• /autoresearch — autonomous hypothesize-experiment-measure loop
• /watch — recurring monitor for new papers, code, or product updates

Skills, Tools, and Compute

Feynman's capabilities ship as Pi skills. Key integrations include AlphaXiv (paper search, Q&A, code reading, annotations via the alpha CLI), Hugging Face Hub (dataset schema and repo file inspection using the public Hub API), web search via Exa, Perplexity, or Gemini API, and session search for indexed recall across prior research sessions. For compute, it supports Docker for isolated local container execution, Modal for serverless GPU burst training and inference, and RunPod for persistent GPU pods with SSH access. Local model providers including LM Studio, LiteLLM, Ollama, and vLLM are supported through the setup flow.

Setup Path

Installation is a single curl or PowerShell command that fetches a standalone native bundle with its own Node.js runtime. A skills-only installer is also available for users who want just the research skill library integrated into Codex, Claude/agent repo setups, or OpenCode projects without the full terminal app. The GitHub README documents version pinning, upgrade, and uninstall paths for all platforms.

Update: v0.2.58

The latest release is v0.2.58, published on May 17, 2026. The repository was created in March 2026 and has seen rapid iteration, accumulating over 7,500 GitHub stars and 928 forks according to the project's own repository metadata. The project signals active development with frequent tagged releases and an open issues count of 7 as of late May 2026.