companion-inc/feynman
Feynman
Feynman is an open-source AI research agent that coordinates multi-agent workflows to search papers, run experiments, and produce cited research briefs. It orchestrates parallel researcher agents that independently investigate subtopics, then synthesizes and verifies findings through a multi-step orchestration loop, enabling deep research across academic papers, web sources, and code.
The tool distinguishes itself through several specialized capabilities, including paper claim verification that audits research paper claims against actual code implementations to identify mismatches and validate results, and a replication planning engine that generates step-by-step plans for reproducing research results with environment selection and compute target configuration. It also functions as a literature review generator, producing structured reviews with cited references covering consensus, disagreements, and open questions on a given topic, and includes a Hugging Face Hub inspector for reading dataset metadata and accessing files from model, dataset, and Space repositories.
Beyond these core differentiators, Feynman supports a broad range of research activities such as paper and web search, cited research brief generation, full-text paper access resolution, paper reading priority ranking, multi-source comparison, ML training recipe discovery, and research paper drafting. It manages research sessions with indexing and search for prior session recall, and can execute bounded experiment loops with benchmark evidence to answer research questions. The tool also provides artifact browsing and preview, system performance observation, and the ability to set up research watches with optional follow-up schedules.
Installation is available via npm into an existing Node.js environment, as a standalone native bundle for macOS, Linux, or Windows with zero external dependencies, or as a research skills-only download for use with Codex or other agents. A guided setup wizard walks through model provider selection, authentication, and optional research-continuity extras like memory or session search.