Companion-Inc/feynman

Feynman – an open‑source AI research assistant (CLI & workbench)

What it is – A command‑line tool (and optional local web workbench) that lets you ask natural‑language research questions and get AI‑generated, source‑grounded answers. It can search papers, the web, code repositories, and a huge range of biomedical and scientific databases, then synthesize briefs, literature reviews, rankings, audits, replication plans, and even ready‑to‑run ML‑training recipes.

How you get it – A one‑line installer downloads a self‑contained native bundle that includes its own pinned Node.js runtime. You can also install the npm package (npm i -g @advaitpaliwal/feynman). A separate installer (install‑skills) lets you add only the “skill library” to other agents (Codex, Claude, OpenCode, etc.) without the full terminal UI.

Key commands

Command	What it does
feynman "<question>"	Quick research brief with citations
feynman rank <topic>	Scores papers on relevance, reproducibility, provenance, etc.
feynman paper <id>	Resolves a DOI/arXiv/PMID and fetches full‑text when possible
feynman lit <topic>	Full literature review with consensus, disagreements, open questions
feynman deepresearch <topic>	Multi‑agent investigation that runs parallel “researcher” agents and then synthesises the results
feynman audit <paper-id>	Checks a paper’s claims against a public codebase
feynman replicate "<paper>"	Plans a reproducibility experiment (runs only after you pick an environment)
feynman recipe "<task>"	Finds ranked, implementable ML‑training recipes from papers and datasets
feynman serve	Starts a local “science workbench” UI with project navigation, chat, notebooks, artifact previews, compute orchestration, provenance tracking, etc.

Built‑in agents – Four specialised sub‑agents are invoked automatically when a workflow needs decomposition:

1. Researcher – gathers evidence from papers, web, repos, and many biomedical databases.
2. Reviewer – critiques the gathered material, grading severity of issues.
3. Writer – turns notes into structured draft text.
4. Verifier – checks citations, resolves dead links, and validates sources.

Data sources & tools – The runtime talks to:

• AlphaXiv for paper search, Q&A and code‑reading.
• A huge suite of life‑science APIs (PubMed, Europe PMC, ChEMBL, UniProt, GWAS Catalog, etc.) for bio‑informatics queries.
• Hugging Face Hub for dataset and model metadata.
• Generic web search providers (with proxy support) and GitHub issue/PR fetchers.
• Local LLM endpoints (LM Studio, Ollama, LiteLLM, vLLM) or hosted providers (OpenRouter, GitHub Copilot, etc.) via feynman model login.

Architecture snapshot

• Pi – the underlying agent runtime that manages packages, extensions, and “skills”.
• Skills – reusable prompt‑plus‑tool bundles (e.g., alpha tools for paper search, Bio Tools for biomedical look‑ups). Skills are stored under ~/.codex/skills/feynman (or repo‑local equivalents) and can be loaded by other agents.
• Workbench – a local Electron‑style UI that stores all sessions, artifacts, and provenance in ~/.feynman/... and optionally syncs to cloud buckets. It includes observability (PostHog, OpenTelemetry) and supports Docker/Modal/RunPod for heavy compute.

Typical workflow

1. Install the CLI (curl … | bash).
2. Run feynman "what do we know about scaling laws" → a cited brief appears.
3. Refine with feynman rank … --expand-citations 2 to build a citation graph.
4. Use feynman deepresearch … for a deeper multi‑agent dive.
5. Open the workbench (feynman serve) to explore artifacts, edit notebooks, or launch a Docker‑based replication run.

Who it’s for – Researchers, data‑scientists, and ML engineers who want a programmable, source‑grounded assistant for literature discovery, reproducibility checks, and rapid prototyping of ML experiments without leaving the terminal (or with the optional UI).

License – MIT (see LICENSE).

Where to learn more – Docs at https://feynman.is/docs, the package stack reference, and release notes linked in the README.