An open-source AI research agent developed by Companion, Inc. Using natural language instructions or slash commands such as /deepresearch and /lit, you can hand off the discovery, reading, comparison, and writing-up of academic papers in one place. Four specialist agents — Researcher, Reviewer, Writer, and Verifier — divide the work, and the design attaches sources to every claim in the output. Rather than a cloud service, it is a CLI tool you install and run on your own machine, licensed under MIT, with the source code published on GitHub.
Key Features
- A four-agent division of labor: The Researcher gathers evidence, the Reviewer critiques the content, the Writer produces drafts, and the Verifier checks citations and sources. Because “the one who investigates” and “the one who confirms” are separate, unsupported assertions are less likely to slip through
- Eight research workflows:
deepresearch(parallel investigation by multiple agents),lit(literature reviews that lay out points of consensus and disagreement),audit(cross-checking a paper’s claims against public code),replicate(planning replication experiments),recipe(finding implementable ML training recipes),rank(scoring papers by citations, method, reproducibility, and provenance),paper(fetching full text from a DOI or arXiv ID), andserve(launching the local workspace) - A local science workbench: Running
feynman servestarts a local application that brings chat, file management, notebooks, compute resources, and audit logs into a single screen, so the trail of an investigation stays in one place - Connections to paper and life-science databases: Alongside AlphaXiv and Hugging Face, it can connect to life-science databases such as PubMed, ChEMBL, and PanglaoDB, as well as web search through Exa and Perplexity
- You choose the model: Local models via LM Studio, Ollama, vLLM, or LiteLLM, plus hosted providers such as GitHub Copilot and OpenRouter, can all be configured through
feynman setup. Running entirely on models you host yourself is possible - Heavy computation offloaded to external GPUs: For replication experiments and benchmarks that need compute, instructions are provided for running through Docker, Modal, or RunPod. No GPU is required for the CLI itself
Pricing
| Plan | Price | Main features |
|---|---|---|
| Open-source edition | $0 (MIT license) | All features available; source published on GitHub |
| Model usage | Depends on the provider you choose | No extra cost with local models; hosted providers such as OpenRouter bill their own API rates |
Pricing reflects information as of August 2026. The software itself is free, but the cost of the AI models used to run it depends on your own provider contracts. Please check the official site for the latest information.
Pros & Cons
✅ Pros
- The software itself is free, and the MIT license imposes few restrictions
- Choosing a local model lets you research without sending unpublished data to an outside service
- Because the design attaches sources to every claim, checking the output is straightforward
- It reaches into steps other literature tools do not cover, such as
auditfor matching claims against public code andreplicatefor building replication plans - Workflows are organized as slash commands, making it easy to start from what you want to do
⚠️ Cons
- As a CLI tool, it suits people comfortable with installation and terminal work; the barrier to entry is higher than browser-based literature services
- You have to prepare and configure the model yourself, and hosted options incur API charges
- Output quality depends heavily on the model you select
- It was released recently, so documentation and worked examples are still accumulating
- Compute-heavy work such as replication experiments requires separately preparing and paying for a GPU environment (Modal, RunPod, and the like)
Comparison with Similar Services
| Item | Feynman | Elicit | Consensus | SciSpace |
|---|---|---|---|---|
| Delivery | Locally run CLI / workbench | Web service | Web service | Web service |
| Main use | Research, review, replication planning, writing | Paper extraction and screening | Finding answers from papers | Reading support and summaries |
| Open source | Published under MIT | Closed | Closed | Closed |
| Model choice | Local or hosted, your pick | Set by the provider | Set by the provider | Set by the provider |
| Code verification | audit cross-checks paper and code | Not supported | Not supported | Not supported |
| Price | Software free (models billed separately) | Free tier / paid plans | Free tier / paid plans | Free tier / paid plans |
Who Is It For
- Researchers and engineers who want to go beyond reading papers and confirm reproducibility and the code behind the claims
- People handling unpublished data or sensitive topics who want their research to stay entirely on a local machine
- Graduate students spending too much time on literature reviews and related-work summaries who want the groundwork automated
- Practitioners who want to turn ML training recipes and replication steps from papers into something implementable
- Anyone who finds the traceability of sources lacking in the output of existing literature-search services
Summary
Feynman is a research agent whose scope extends past literature search into verifying a paper’s claims and planning replication. It does assume CLI operation and that you supply your own model, but in exchange the software itself is free and you retain the option of keeping research entirely local. A good starting point is to configure a model with feynman setup and try lit or deepresearch on a topic of your own to see how much you can hand over.