Yes, if your goal is to embed research skills into an existing agent rather than use the full terminal application or science workbench.

• The README explicitly says the skills-only installer does not install the Feynman terminal, bundled Node runtime, authentication storage, or Pi packages.
• Codex skills can be installed under ~/.codex/skills/feynman; repository-local Claude/agent skills go under .agents/skills/feynman.
• OpenCode has a separate .opencode/skills/feynman target for project-local installation.
• The installation includes the bundled skills/ and prompts/ trees plus referenced repository guidance files. This reuses research capabilities but is not equivalent to the full session, workbench, Notebook, compute, or file-management experience.

If you need feynman serve, workbench state, or the complete Pi runtime, skills-only is insufficient.

Yes. Prefer the standalone native bundle because it includes a pinned Node.js runtime and supports explicit release pinning; npm installation depends on the machine’s local Node.js.

• On macOS/Linux, the installer downloads a standalone native bundle with its own pinned Node.js runtime.
• Before replacing an existing installation, it verifies the release SHA-256, which supports deployment processes that require package-integrity checks.
• The README allows explicit version pinning, for example curl -fsSL https://feynman.is/install | bash -s -- 0.2.35.
• The npm alternative uses the local Node.js runtime. Also, feynman update refreshes only Pi packages and does not replace the standalone runtime bundle, so upgrade responsibilities differ.

Choose standalone when runtime consistency matters. npm is more natural only when the team already centrally manages Node.js and npm-based distribution.

Yes. It is suitable because it combines paper analysis, paper-to-code auditing, replication planning, and ML training-recipe ranking in one research-agent environment.

• /audit is explicitly described as a “Paper vs. codebase mismatch audit,” so it can compare claims with implementation details.
• /replicate plans replication checks, but execution happens only after an environment is chosen; it is not an unconditional one-click reproduction system.
• /recipe ranks ML training recipes by dataset, method, code, and verification status, which fits candidate-method screening.
• Paper resolution accepts DOI, arXiv ID, OpenAlex ID, PMID, PMCID, or title, with OpenAlex, arXiv/alphaXiv, DOI, and Europe PMC sources.

You still need to verify whether the target repository’s language, dependencies, hardware, and execution model are supported by the available compute and Files environment.

It depends. Feynman explicitly provides a bounded experiment loop and compute-related workbench capabilities, but the README does not guarantee that arbitrary projects execute out of the box or provide default resource, permission, and stopping policies.

• /autoresearch is described as a “Bounded experiment loop with benchmark evidence,” directly matching benchmark-driven iteration.
• feynman serve includes compute, Files host inventory, and local, SSH/BYOC, and cloud-backed artifact contexts.
• The workbench ledger includes compute egress, Modal environment fields, and compute poller lease state, indicating awareness of compute-task state management.
• Project insights warn that autoresearch, compute, and code execution may consume resources, pollute environments, or cause side effects. The stated limitations also tie execution to dependencies, hardware, containers, or remote-environment configuration.

It fits an ML workflow with a controlled execution environment, but not treating an arbitrary repository as a safe autonomous experimentation platform.

It depends. Model inference can use local services, but literature search and source retrieval may still require external network access, so local inference alone does not guarantee a fully offline workflow.

• The README explicitly supports LM Studio with the default endpoint http://localhost:1234/v1; LiteLLM uses http://localhost:4000/v1 by default.
• Ollama and vLLM can be configured through Custom provider, openai-completions, and a local /v1 endpoint.
• /lit, /review, and /draft cover literature review, research review, and paper-style drafting from findings.
• However, the project uses OpenAlex, arXiv/alphaXiv, DOI, and Europe PMC sources; the README does not promise that these work without network access.

If all material, credentials, and sources must remain inside an isolated network, applicability depends on whether you can provide local source copies and a supported access path.

It depends. The feature coverage matches long-running research-record needs, but the project’s release data and stated limitations indicate that some workbench and scientific-tool capabilities may still be evolving quickly.

• feynman serve provides project/session navigation, notebooks, compute, Files host inventory, sources, settings, lineage, and provenance.
• It also supports previews for audio/video, spreadsheets, notebooks, LaTeX, and scientific artifacts, plus artifact Notes and note-preview modals.
• Workbench state is stored under ~/.feynman/orgs//workbench, with a feynman-workbench.db mirror, which fits structured research records.
• However, the latest release is v0.3.48 and the project has 5 releases; the README does not promise upgrade compatibility, backup/restore, multi-user collaboration, or data migration.

It is therefore suitable for individual or small-scope continuing research, but the available information is insufficient for mature team governance and defined data lifecycle requirements.