Feynman AI Research Agent

The Science Workbench provides a comprehensive suite of research tools:

• Projects and Sessions: Manage research projects, open existing sessions, and maintain Pi-backed chat metadata.
• Feynman Bio Tools: Access a wide array of science connectors for literature (OpenAlex, arXiv, PubMed), chemistry (PubChem, ChEBI, Ketcher), genomics (Ensembl, MyGene, UniProt), and various omics/pathway databases.
• Artifacts and Previews: Browse and preview diverse file types including Jupyter notebooks, LaTeX, chemistry artifacts (SMILES, SDF, etc.), proteins, alignments, and tensors. Supports HTML report annotation with selector/text capture.
• Notebooks and Compute: Execute Python, R, and Bash cells, inspect session kernels, and manage compute providers and jobs.
• Versions and Lineage: Inspect artifact versions, checksums, producer records, and execution evidence.
• Settings and Resources: Manage specialists, skills, permissions, compute, network, storage, and credentials (viewing redacted availability without exposing raw secrets).

Use the managed-model-endpoints skill to make a model endpoint usable within Feynman. This is intended for research workflows that require local or remote model services, endpoint health checks, credential references, startup scripts, or inference routing.

Implementation Workflow

1. Define Endpoint Specifications: Determine the endpoint purpose, model family, input/output schema, authentication method, hardware requirements, and expected latency.
2. Identify Endpoint Type: Categorize the endpoint as local, remote HTTP, Modal-backed, SSH-backed, or a custom connector.
3. Configure Metadata and Secrets:
   • Add non-secret endpoint metadata directly to settings.
   • Do not store secrets in settings. Store secret references as environment variable names or credential references.
4. Lifecycle Management: If Feynman is responsible for the endpoint lifecycle, implement the following checks:
   • start
   • stop
   • health
   • inference
5. Validation: Perform a small inference smoke test and save the request/response shape, ensuring no secrets are leaked in the saved data.

Best Practices

Expose endpoints as research infrastructure. Do not treat an endpoint as a permanent installation claim until health checks have successfully passed.

If you only want Feynman's research skills (the skills/ and prompts/ trees) without the full terminal runtime, bundled Node runtime, or auth storage, use the skills installer.

macOS or Linux

• Codex install (~/.codex/skills/feynman):

  curl -fsSL https://feynman.is/install-skills | bash
  # Or explicitly:
  curl -fsSL https://feynman.is/install-skills | bash -s -- --codex

• Repo-local Claude/agent install (.agents/skills/feynman):

  curl -fsSL https://feynman.is/install-skills | bash -s -- --repo

• OpenCode project-local install (.opencode/skills/feynman):

  curl -fsSL https://feynman.is/install-skills | bash -s -- --opencode

Windows (PowerShell)

• Codex install:

  irm https://feynman.is/install-skills.ps1 | iex
  # Or explicitly:
  & ([scriptblock]::Create((irm https://feynman.is/install-skills.ps1))) -Scope Codex

• Repo-local install:

  & ([scriptblock]::Create((irm https://feynman.is/install-skills.ps1))) -Scope Repo

• OpenCode install:

  & ([scriptblock]::Create((irm https://feynman.is/install-skills.ps1))) -Scope OpenCode

The indication-dossier skill is used to assemble a source-backed biomedical evidence packet for a disease, target, mechanism, biomarker, or therapy. This is ideal for research tasks involving disease biology, target rationale, patient segmentation, biomarkers, clinical trials, drugs, competitive landscape, or translational evidence.

Workflow

1. Define Scope: Specify the indication, patient segment, intervention or target, and the specific decision the dossier must support.
2. Gather Evidence: Collect data from literature, clinical trials, Open Targets, ChEMBL, CIViC, ClinGen, cBioPortal, DepMap/COSMIC-compatible sources, and other Feynman Bio Tools.
3. Categorize Evidence: Organize findings into distinct sections: human evidence, model-system evidence, mechanism, biomarkers, clinical precedent, safety, and open questions.
4. Evaluate Claims: Rank claims based on source strength and reproducibility. Explicitly flag missing cohorts, potential confounders, and assay limitations.
5. Export: Save the completed dossier and its accompanying provenance sidecar to the outputs/ directory.

Constraint: Do not convert sparse evidence into clinical recommendations.

Use the openfold3 skill when a task requires OpenFold-family structure prediction, setup, model comparison, or ensuring reproducibility for OpenFold-family outputs.

Follow this workflow for successful execution:

1. Normalize Inputs: Ensure FASTA, chains, templates, MSAs, ligands/partners, seeds, and the expected output format are normalized.
2. Verify Environment: Before execution, verify the model code, checkpoints, databases, and the GPU route.
3. Capture Provenance: Save the input manifest, environment lockfile, command, logs, structure outputs, confidence files, and runtime metadata.
4. Compare Results: When results affect decisions, compare them against AlphaFold-style, ESMFold-style, PDB, or literature evidence.
5. Report Gaps: Explicitly report reproducibility gaps such as missing databases, unavailable weights, failed templates, or route-specific approximations.

Important: Never present an OpenFold-family output without specifying the version and input provenance.