Turn Any AI Agent Into a Scientist with Scientific Agent Skills

Most coding agents stop at writing and refactoring code. Scientific Agent Skills hands them 161 ready-to-use skills that reach into biology, chemistry, medicine, and materials science — querying PubChem, running RDKit, docking molecules against AlphaFold structures — so a research pipeline becomes a single conversation.

Why This Skill Matters

Scientific computing is scattered. A typical project pulls from dozens of public databases (ChEMBL, UniProt, ClinicalTrials.gov, PubChem), each with its own API and pagination quirks, and from specialized Python packages (RDKit, Scanpy, PyTorch Lightning, scikit-learn), each with version-sensitive APIs. You normally spend your time pasting docs into the agent and correcting hallucinated function calls.

This collection removes that friction. Each skill encodes the exact, version-aware workflow for one database or package: which endpoint to hit, which arguments are required, how to parse the result, and how to chain it with the next step. The agent reasons over the workflow instead of guessing. The skills run on Cursor, Claude Code, Codex, and Gemini CLI, so one install follows you across hosts.

Installation

npx skills add K-Dense-AI/scientific-agent-skills

That pulls the whole collection into your agent's skills directory. If you prefer the GitHub CLI:

gh skill install K-Dense-AI/scientific-agent-skills

The full set is large. Install only the topical subset you need (see Tips) to keep your agent's context window lean.

Real Workflow: A Drug-Discovery Pipeline

A realistic scenario, not "hello world": find candidate kinase inhibitors, analyze them, and check the literature for resistance.

The pipeline:

1. Query ChEMBL for inhibitors of a target protein.
2. Use RDKit to analyze structure-activity relationships and drug-likeness.
3. Dock the top candidates against an AlphaFold structure with DiffDock.
4. Search PubMed for reported resistance mutations.
5. Assemble a visual report.

Prompt your agent:

Find ChEMBL inhibitors of EGFR, compute drug-likeness with RDKit,
dock the top three against AlphaFold's EGFR structure using DiffDock,
and summarize PubMed-reported resistance mutations. Output a markdown report.

Because each skill encodes the correct calls and argument shapes, the agent walks the steps without you babysitting API details. Database skills need an internet connection; the Python-package skills run offline once installed.

The agent returns a structured markdown report you can hand to a colleague — for example:

## EGFR Inhibitor Screen — Draft Report

| Compound | ChEMBL ID | Drug-likeness | Docking score |
|---|---|---|---|
| <inhibitor> | CHEMBL... | <Lipinski violations> | <kcal/mol> |

PubMed resistance: <mutations and which inhibitors they affect>

The exact compounds and scores depend on your query and installed library versions; treat the report as a draft for human review.

Real Workflow: A Quick PubChem Lookup

For smaller tasks, a single skill is enough. Search a compound and pull its properties:

Look up caffeine in PubChem and return its molecular weight, formula,
and a canonical SMILES string.

The PubChem skill knows the right REST endpoint and how to parse the JSON, so you get structured fields instead of a pasted webpage.

For caffeine, the skill returns structured fields straight from the REST API:

name:    caffeine
formula: C8H10N4O2
weight:  194.19 g/mol
SMILES:  CN1C=NC2=C1C(=O)N(C(=O)N2C)C

Tips

• Install a subset first. All 161 skills at once can crowd your context. Add one domain (cheminformatics, bioinformatics, clinical research) and expand as needed.
• Treat skills like dependencies. A skill can execute code and steer agent behavior, so read anything you install — the same caution you'd apply to an npm package.
• Mind the versions. Scientific Python libraries change fast; the skills are version-aware, so pin and note the versions in your environment to avoid drift.
• Run the bundled tests. Every skill ships documentation, examples, and a test suite. Run them after install to confirm the skill behaves in your setup.

When Not to Use This

• Clinical diagnosis or treatment decisions. The clinical skills are for retrospective validation only — not patient-specific diagnosis, alarms, or deployment. Treat every clinical output as a draft for qualified human review.
• Fully offline database work. Database skills call live APIs over the internet. In an air-gapped environment, only the Python-package skills are usable.
• Tasks outside these domains. If your task never touches a scientific database or package, the collection adds context weight for no benefit.

See the leaderboard for more skills.