Scientific Agent Skills

Scientific Agent Skills is an MIT-licensed, open-source collection of 168 procedural knowledge skills for AI research agents, created by K-Dense Inc. and described in an arXiv paper (arXiv:2609.00065) submitted in August 2026. The library is compatible with any agent host that supports the open Agent Skills standard, including Cursor, Claude Code, Codex, Gemini CLI, and Google Antigravity. According to the GitHub repository description, it is used by 250,000+ scientists worldwide and has accumulated over 47,000 stars.

What It Is

Scientific Agent Skills is a structured skill registry — a GitHub repository where each skill is a directory containing a versioned, human-readable SKILL.md instruction file, optional reference materials, and runnable scripts. When an AI agent encounters a task that matches a skill, it loads the relevant file to get curated, field-validated procedural guidance: which statistical test the field accepts, which identifier namespace is authoritative, and which caveats must accompany a result. The library is also a valid Agent Plugins 1.0.0 package, so plugin-capable clients can load the entire collection as a single plugin.

Skill Coverage and Domain Breadth

The 168 skills span 20+ scientific domains organized into named categories:

• Bioinformatics & Genomics (28 skills) — Scanpy, PyDESeq2, scVelo, TileDB-VCF, AlphaGenome Atlas lookups, OneKGPd population genomics, Genomic Intelligence hosted predictions, and pathogen surveillance via GenSpectrum LAPIS
• Cheminformatics & Drug Discovery (10 skills) — RDKit, DeepChem, DiffDock, OpenMM/MDAnalysis molecular dynamics, Rowan cloud quantum chemistry
• Clinical Research & Evidence Workflows (8 skills) — PK/PD modeling (NCA, population PK, bioequivalence, first-in-human dose), DepMap, Imaging Data Commons, clinical decision-support research artifacts
• Machine Learning & AI (14 skills) — PyTorch Lightning, Transformers, scikit-learn, TimesFM, PyMC, Torch Geometric, SHAP
• Scientific Communication (27 skills) — Paperclip full-text corpus (~11M papers, 217K+ regulatory documents), Paper Lookup across 10 academic databases, evidence-traceable Scientific Writing, macro-free PPTX posters, Mermaid diagrams
• Scientific Databases (12 skills → 100+ databases) — A unified database-lookup skill covers 78 public databases (PubChem, ChEMBL, UniProt, ClinVar, ClinicalTrials.gov, FRED, USPTO, and more); dedicated skills add DepMap, PrimeKG, NCATS ARAX, AlphaGenome, and others
• Regulatory & Standards (2 skills) — ISO 13485, ISO 14971, ISO/IEC 17025, ISO 15189 evidence-preparation artifacts; ICH Q2(R2)/Q14, ICH M10, USP, CLSI analytical method validation

Installation and Compatibility

Skills can be installed via three primary paths. The npx skills add K-Dense-AI/scientific-agent-skills command works on supported hosts including Claude Code, Codex, Gemini CLI, and Cursor. The GitHub CLI (gh skill install) supports interactive browsing, per-skill installation, version pinning to a release tag or commit SHA, and automatic provenance metadata recording. For Agent Plugins-compatible clients (Cursor, Codex, GitHub Copilot, VS Code, Kiro), the repository can be symlinked or copied into the local plugins directory. Manual git clone into ~/.agents/skills/ or .agents/skills/ covers additional hosts such as OpenClaw, NemoClaw, Pi, and Hermes.

Architecture and Safety Design

Each skill directory is built around a SKILL.md with YAML frontmatter, a metadata.version field, and explicit safety boundaries. The paper reports that the always-resident descriptions of all 163 skills (at time of writing) cost 7.1% of a 200,000-token context window, and the median documented workflow fits within 23.9% of it. Skills are loaded on demand rather than kept in standing context, which keeps token costs manageable. The repository runs LLM-based security scans via the Cisco AI Defense Skill Scanner on every skill weekly, with full rescans at least every 30 days, and publishes results to docs/security-report.md. A CI structural contract enforces frontmatter conformance, link resolution, script parsing, and --help behavior on every pull request; skills that ship scripts/ must include a test suite under tests/<skill-name>/.

Update: v2.70.0

The latest release is v2.70.0, published on September 29, 2026. The arXiv paper was submitted August 30, 2026 (v1) and revised September 2, 2026 (v2), which added a figure of how documented workflows compose skills, quoted skill clauses behind the introduction's examples, named supported hosts and the pinned install path, and reported two corpus findings in the abstract. The repository was originally launched as "Claude Scientific Skills" and rebranded to "Scientific Agent Skills" to reflect compatibility with any Agent Skills-standard host, not just Claude. The skill count has grown from 163 at paper submission to 168 in the current release.