Skip to content

Note feynman

All posts
Nov 13, 20269 min readfeynman · ai-agents · research · open-source · arxiv · literature-reviewRead in:ENID

Feynman: The Open Source AI Research Agent That Automates Literature Reviews

Feynman launched in March 2026 and already has 10,117 stars on GitHub. Four AI subagents search arXiv, PubMed, and Semantic Scholar, verify every citation, and can audit a paper's claims against its public code.

A researcher working on a laptop among books and papers in a library. Photo by Ron Lach on Pexels.

A literature review can eat a whole week. You open one paper, follow its citations to five more, check whether the claims hold up, and by Friday you have a folder of PDFs and a half written summary. Feynman is an open source CLI agent built to take over that grind. It searches the major paper databases, sends four AI subagents to read, verify, and critique what it finds, and writes drafts with citations it actually checks. It launched in March 2026 and has already passed 10,000 stars on GitHub. (GitHub)

This post explains what Feynman is, how its four agents divide the work, what its ten workflows can do, and how to install it. No research background needed.

Key Takeaways

Feynman is a free, open source CLI research agent from Companion, Inc., MIT licensed, with 10,117 stars and 1,108 forks on GitHub as of November 13, 2026.

Four subagents share the work: Researcher finds papers, Verifier checks citations and claims, Reviewer critiques drafts, and Writer turns findings into text.

It searches arXiv, PubMed, Semantic Scholar, OpenAlex, bioRxiv, medRxiv, Crossref, and your local PDFs, and it can audit a paper's claims against its public code.

Ten workflows cover deep research, literature reviews, paper reviews, replication plans, training recipes, and paper drafts.

Install takes one command on macOS or Linux, and it can run fully locally with Ollama or any OpenAI compatible local model.

What Is Feynman?

Feynman is a command line tool that turns a research question into a structured, cited answer. You type a question, and instead of one chatbot reply, you get a small team of agents working through it. The project describes itself as "the open source AI research agent", and its tagline is that it reads papers, searches the web, writes research drafts, plans experiments, and cites research claims. (Feynman)

It is built on two foundations: Pi, an open source coding agent runtime, and alphaXiv, a paper search and analysis service. That combination matters because Feynman is not a chatbot with a search plugin. It is an agent system with tools, file outputs, and a verification step built into the pipeline.

Watch the 30-second repo signal on Repomaniac to see it in action.

Four Agents, One Job Each

The core idea is simple: one agent doing everything makes mistakes, so Feynman splits research into four roles that check each other.

Aspect

Detail

Researcher

Finds and reads papers across the literature databases

Verifier

Checks citations, removes dead links, flags unsupported claims

Reviewer

Critiques drafts with severity grades: critical, major, minor, nit

Writer

Turns verified findings into drafts, marking unsubstantiated content as TODO

The Researcher does not run one search and stop. For each question it runs 2 to 4 reworded queries, merges the results, and routes each query to the right source: general machine learning and computer science papers go through Semantic Scholar, biomedical questions go through PubMed and then Europe PMC, and citation graphs go through OpenAlex. Its evidence file must hold at least 5 entries, each with a stable numeric source ID, so nothing in the final draft traces back to a single paper. (Feynman Docs)

The Verifier is the part most AI tools skip. It checks every citation in a draft, removes dead links, and flags claims the sources do not support. The Reviewer then grades problems by severity, from critical down to nit, so you know what to fix first. The Writer marks anything unsubstantiated as a placeholder instead of inventing a citation, which is exactly the failure mode that makes researchers distrust AI summaries. (AcademAI)

The 10 Workflows

Feynman ships with ten slash workflows, each one a full research task rather than a single prompt. You run them from the terminal, for example `feynman lit "RLHF alternatives"` or `feynman deepresearch "mechanistic interpretability"`.

Aspect

Detail

/deepresearch

Planned investigation with parallel researchers and citation checks

/lit

Literature review: consensus, disagreements, open questions

/review

Critique a paper or draft with severity levels and a revision plan

/audit

Paper claims compared with its public code

/replicate

Replication plan, runs only after you choose an environment

/recipe

Ranked ML training recipes with datasets and code

/compare

Source comparison matrix

/draft

Paper-style draft from research findings

/autoresearch

Bounded experiment loop: hypothesize, experiment, measure, repeat

/summarize

Summary of a paper, repository, or PDF

Deep research runs in five phases: planning, agent dispatch, data extraction, synthesis, and citation verification. Results are saved to an `outputs/` folder, drafts go to `papers/`, and the literature review, deep research, and recipe workflows also write a provenance file listing the sources used and what was verified. (GitHub, AcademAI)

The most unusual workflow is `/audit`. Point it at a paper and it compares the paper's claims against the paper's own public code repository. A typical call looks like `feynman audit 2401.12345`, using the paper's arXiv ID. For anyone who has wondered whether a paper's results actually match the code it shipped, that is a genuinely useful check. (VirtualUncle)

Where Feynman Finds Papers

Feynman searches the databases researchers already use, not a scraped copy of the web. The official list covers Semantic Scholar, OpenAlex, arXiv, PubMed, Europe PMC, bioRxiv, medRxiv, and Crossref, plus Hugging Face, general web research, and local PDFs and office documents on your own machine. (Feynman)

That range covers both sides of research: computer science and machine learning through arXiv and Semantic Scholar, and life sciences through PubMed, Europe PMC, and the preprint servers bioRxiv and medRxiv. If you already keep a folder of PDFs, Feynman reads those too, so a literature review can include papers that are not in any public index.

Free API keys for OpenAlex and Semantic Scholar are recommended to avoid shared rate limits, but the tool works without them. Heavy compute, like running replication experiments, can go through Docker locally or through Modal and RunPod for cloud GPUs. Local models are supported through LM Studio, LiteLLM, Ollama, and vLLM. (EveryDev)

How to Install Feynman

Installation is one command, and the project is under active development: the latest release, v0.5.68, shipped on November 11, 2026, two days before this post. (GitHub)

On macOS or Linux:

``` curl -fsSL https://feynman.is/install | bash ```

On Windows, in PowerShell:

``` irm https://feynman.is/install.ps1 | iex ```

Or with npm, if you have Node.js 22.22 or newer:

``` npm install -g @companion-ai/feynman ```

Then sign in to a model provider:

``` feynman setup ```

From there, a first question is as simple as `feynman "what do we know about scaling laws"`. The installation guide also covers pinned versions, uninstalling, and a skills-only bundle for Codex and OpenCode, and the setup guide covers local models and editors such as Zed.

Privacy: Your Papers Stay on Your Machine

Feynman is an orchestration layer, not a hosted service. There is no Companion server in the middle: session history and tokens are stored locally under `~/.feynman/`, and a fully local mode is possible with Ollama or any OpenAI compatible local server. (AcademAI)

One tradeoff to know about: Feynman sends anonymous usage telemetry to PostHog, covering commands, workflows, tools, models, token counts, and errors with their messages. It never sends prompts, model output, paper content, or tool arguments, and you can turn it off entirely with `FEYNMAN_TELEMETRY=off`. (GitHub)

Feynman vs a General Chatbot

The difference is not the model, it is the pipeline. A general chatbot answers from memory and may cite a paper it never opened. Feynman searches real databases, assigns a verifier to check the citations, and writes the provenance down.

Aspect

General chatbot

Feynman

Sources

Training data, web search

arXiv, PubMed, Semantic Scholar, OpenAlex, local PDFs

Citations

Often invented or unchecked

Verified by a dedicated agent, dead links removed

Output

One reply in chat

Files in outputs/ and papers/, plus a provenance file

Paper vs code check

Not built in

/audit compares claims against public code

Local mode

Usually cloud only

Ollama and local models supported

That said, Feynman is a research assistant, not a replacement for reading. The Verifier catches unsupported claims, but it cannot judge whether a paper's method is sound or whether a result matters for your field. The sensible workflow is Feynman for the first pass, then you read the papers it flags as most important.

Why It Took Off So Fast

Feynman was created on March 19, 2026, and it passed 2,300 stars within days of launch. By May it was past 7,500 stars, and by November 13 it stood at 10,117 stars with 1,108 forks and 168 releases. That pace, roughly 10,000 stars in under eight months, puts it among the fastest growing research tools of the year. (GitHub, VirtualUncle, EveryDev)

Part of the appeal is timing. AI coding agents are now common, but research agents that verify their own citations are still rare. Feynman also arrived as an open source project under the MIT license, so researchers can read the code, self host it, and extend it with their own tools, for example by adding a Zotero library through `feynman mcp add`.

Early users report real gains in practice. One author who wired Feynman skills into a Claude based second brain system measured Hit@1 on their own action search problem rising from 27.6 percent to 57.3 percent. That is the author's own benchmark on their own problem, not a Feynman product metric, but it shows the kind of workflow people are building around it. (StackOne)

Frequently Asked Questions

Is Feynman really free?

Yes. Feynman is open source under the MIT license, and the tool itself costs nothing. You still need access to a language model, either a paid provider account or a local model through Ollama or a similar server.

Do I need to know how to code to use Feynman?

You need basic terminal comfort: installing it, typing a question in quotes, and reading files in the outputs folder. You do not need to write code. The workflows are plain language commands like `feynman lit "your topic"`.

Can Feynman read my own PDFs?

Yes. It reads local PDFs and office documents, so you can point it at papers you already downloaded and include them in a literature review alongside the public databases.

How is Feynman different from a search engine?

A search engine returns links. Feynman reads the papers, compares their claims, checks the citations, and writes a structured draft with a provenance file, so you can trace every claim back to its source.

Does Feynman work offline?

Partly. With a local model through Ollama, the agent itself runs locally, but searching arXiv, PubMed, and the other databases still needs an internet connection. Local PDFs can be read without it.

Conclusion

Feynman does one thing well: it turns the slow, mechanical part of research into a pipeline you can run from the terminal. Four agents divide the work, a verifier checks the citations, and the audit workflow tests papers against their own code. It is free, open source, actively developed, and it can run entirely on your machine.

If you read papers for a living, or you are starting a literature review and dreading the first fifty PDFs, install it and ask it one question. The first literature review it writes for you will tell you more than any article can.

Baca versi Bahasa Indonesia dari artikel ini di Feynman: Agen Riset AI Open Source yang Mengotomatiskan Tinjauan Literatur.

Tags: feynman · ai-agents · research · open-source · arxiv · literature-review

All posts