No description
Pre-1.0 at v0.2.54, but 54 patch releases in roughly two months (March 19 → May 11, 2026) signals extremely rapid iteration rather than stability. The 0-open-issues count with active development suggests disciplined triage. No stable API guarantee is implied at this version band, and the README itself notes the feynman update / standalone-bundle distinction — a sign the install model is still settling. Promising but not yet production-stable.
README is dense and practical: platform-specific install instructions (macOS/Linux/Windows), version-pinning guidance, uninstall steps, local-model setup (LM Studio, LiteLLM, Ollama, vLLM), a full workflow table, agent roster, skills/tools inventory, and architecture narrative. A dedicated docs site exists at feynman.is/docs. CONTRIBUTING.md and RELEASES.md are shipped in the package. Gap: no in-repo API reference or skill authoring guide visible from the README alone; that material presumably lives on the external site.
TypeScript throughout with a strict build config (tsconfig.build.json) and separate typecheck script. Test suite present (node --test, concurrency-limited to avoid flakiness). Dependency overrides for known CVE-adjacent packages (path-to-regexp, express-rate-limit, ip-address, fast-uri) show active hygiene. Uses TypeScript 6.0.3 — leading edge as of mid-2026. tsx for dev, tsc for production dist is a clean pattern. No CI badge visible in README, which is a minor gap.
Last commit and latest release both dated May 11, 2026 — the day before appraisal. Roughly one release per calendar day since inception. Zero open issues. The company (companion-inc) appears to be the primary maintainer, suggesting organizational backing rather than solo-author bus-factor risk.
7,062 stars and 883 forks in ~53 days is a steep trajectory — comparable to breakout CLI tools at launch. Fork ratio (~12 %) is healthy, suggesting genuine downstream use rather than passive starring. No visible downstream npm dependents yet, consistent with the tool's age. Star velocity is the strongest signal here.
Overall: 3.9/5
Category: AI Research Agent CLI
Known alternatives in vault: karpathy--autoresearch (Autonomous ML Research, 3/5); VoltAgent--awesome-ai-agent-papers (AI Agent Paper Curation, 3.4/5)
Differentiation: Feynman is substantially more capable than karpathy--autoresearch: it ships a full multi-agent dispatch layer (Researcher / Reviewer / Writer / Verifier), experiment replication on real GPU compute (Modal, RunPod), paper-vs-codebase audit (/audit), ML training recipe synthesis (/recipe), and a source-grounded citation pipeline. It also provides a skills-only install path that drops directly into Codex, Claude, and OpenCode agent environments — a deployment surface neither vault alternative addresses. karpathy--autoresearch appears to be a lighter autonomous loop; VoltAgent--awesome-ai-agent-papers is a curated list, not an agent. What alternatives may do better: karpathy--autoresearch is presumably simpler to audit/trust at the source level; Feynman's dependency on Pi, alphaXiv, and external services introduces more moving parts.
Gap or crowd: The "Autonomous ML Research" category has only one prior entry (thin coverage). Feynman is meaningfully more featureful, arguably warranting its own sub-category ("AI Research Agent CLI") rather than being lumped with lighter autonomous-loop tools.
Score: 5/5
Harvestable: (1) Skills library — literature review, deep research, paper audit, recipe generation, and auto-research prompts are installable as agent skills via a dedicated one-liner (install-skills). (2) Multi-agent dispatch pattern (Researcher/Reviewer/Writer/Verifier) is a directly reusable workflow template. (3) Source-grounding methodology — every claim linked to a paper, doc, or repo URL — is a harvestable epistemic pattern for vault ingestion pipelines. (4) AlphaXiv + Hugging Face Hub tool integrations are concrete examples of research-data-source wiring. (5) Session-indexed recall across prior research sessions is directly analogous to personal knowledge vault retrieval.
Integration path: The install-skills script supports --repo (drops into .agents/skills/feynman) and --codex/--opencode targets — meaning skills slot into a Claude/Codex/OpenCode agent environment with a single command, no custom wiring. The full CLI can additionally be invoked as a subprocess tool from any orchestrator. Research outputs are source-grounded text artifacts, well-suited for direct vault ingest.
Overlap with existing: karpathy--autoresearch overlaps on autonomous ML research execution. mattpocock--skills overlaps on the skills-library deployment pattern (different domain: coding vs. research). VoltAgent--awesome-ai-agent-papers overlaps marginally on paper curation. None of these replicate Feynman's combined agent dispatch + GPU compute + alphaXiv integration.
Adoption cost: Trivial for skills-only use (one installer command, no auth required beyond optional HF_TOKEN). Moderate for full CLI integration: requires Feynman account or local-model setup, Modal/RunPod credentials for replication workflows, and alphaXiv account for full paper-Q&A depth.
Feynman is the most directly PAI-relevant repo appraised to date: it ships a research agent with a dedicated skill-installation path designed explicitly for agent runtimes like Codex and Claude. The skills-only install surface means the vault can harvest Feynman's research workflows without adopting the full CLI stack. The rapid release cadence (v0.2.54 in 53 days) and zero open issues suggest a team actively eating their own dog food. Primary risk is the pre-1.0 API instability and the dependency chain on Pi (badlogic/pi-mono) and alphaXiv — both external projects whose own stability and availability could affect Feynman's reliability. Recommend tracking: watch for v1.0 stabilization signal, and evaluate extracting /lit and /recipe skill prompts into the vault's own prompt library even if the full CLI is not adopted.