Feynman combines AI research with auditable research artifacts

What this is about

Feynman is an open-source research agent from Companion that does not stop scientific research at a chat answer. The tool searches academic literature and web sources, creates cited briefs, compares claims in papers with public code, and plans replication work. It runs in a terminal or as a local browser-based science workbench.

The product is mainly relevant to researchers, students, and technical teams that want to inspect results, sources, and intermediate steps together. Its official documentation describes installation on macOS, Linux, and Windows, plus a separate skills package for Codex, Claude, and OpenCode.

What Feynman actually does

The command-line tool accepts a research question and routes it according to the task. feynman lit creates a literature review with consensus, disagreements, and open questions. feynman rank orders papers using evidence about citations, methods, reproducibility, and provenance. feynman audit compares a paper’s claims with its publicly available code. feynman replicate drafts a replication plan; experiments are meant to run only after the user explicitly selects an execution environment.

Running feynman serve opens a local workbench for projects, chats, files, notebooks, artifacts, and provenance. For literature and scientific metadata, the project integrates sources including alphaXiv, OpenAlex, arXiv, PubMed, and Europe PMC. Local models can be connected through LM Studio, LiteLLM, Ollama, or vLLM, while hosted model providers are also supported.

Installation is available as a standalone runtime bundle or through npm. The standalone installer downloads a tagged release, verifies its SHA-256 checksum, and includes a compatible Node.js runtime. Users who only need the research workflows can install the skills separately.

Why it matters

General chatbots can produce a literature list, but the real work starts afterward: Which source supports which claim? Is the code public? Can a result be repeated under realistic conditions? Feynman attempts to combine these steps in an auditable workflow. Its value lies less in one answer than in the link between research, citations, artifacts, and history.

That matters for teams when a report must later be reviewed or continued. Instead of managing loose chat transcripts, PDFs, tables, and notebooks separately, the local workbench keeps these elements in one project context. This does not automatically reduce errors, but it can make mistakes and missing evidence easier to find.

Feynman is available under an open license on GitHub. The product page identifies Companion as the provider, while the repository and documentation show the installation path and supported research functions. The tool is therefore directly testable rather than merely announced on a landing page.

In plain language

Feynman is like a carefully organized travel folder. A normal chatbot may write a list of sights. Feynman also tries to store tickets, opening hours, maps, receipts, and changes so that another person can reconstruct the trip. The folder may still contain a wrong timetable, but the origin of the plan is more visible.

A practical example

A research team wants to test whether a training method reduces the error rate of a small language model. It first asks Feynman to collect 40 relevant papers and rank them by methodology, citation graph, and reproducibility. The team then selects five papers for full-text review.

Using the audit command, it compares one paper’s central claims with the accompanying repository. It then creates a replication plan for three training runs with the same dataset and a fixed compute budget. Results, configurations, tables, and a notebook remain in the same local workbench. Only the researchers decide whether to execute the proposed experiments in Docker, on their own hardware, or with a compute provider.

The sensible first test is smaller: install Feynman, ask one narrowly scoped question, and manually verify ten retrieved sources. The team should then check whether citations, versions, and conclusions actually match the originals.

Scope and limits

First, the quality of search results and synthesis depends on the selected model, reachable databases, and available full text. A neatly formatted source list does not prove that the interpretation is correct.

Second, a local workbench does not automatically provide complete data sovereignty. When users choose a hosted model provider or external search service, prompts are sent to those services. Sensitive research data should enter the workflow only after reviewing configuration, contracts, and logging.

Third, Feynman replaces neither peer review nor statistical and domain expertise. Automatically generated rankings can miss important niche work, and a replication plan can contain methodological errors. Model access, compute, and data costs also depend on the selected environment; no universal total price is stated.

SEO & GEO keywords

Feynman, Companion, AI research agent, Open Source AI, literature review, scientific research, PaperRank, reproducibility, local science workbench, research artifacts, alphaXiv, Codex Skills