Ask your AI coding agent to dock a molecule against a protein, and it will happily write the code. Ask it which protonation state the ligand should carry at physiological pH, and you get a confident answer that may be wrong in a way that invalidates the whole run. That gap, between being able to call a library and knowing how a field actually uses it, is what Scientific Agent Skills exists to close.
It is a free, MIT-licensed library of 177 validated skills plus 100+ scientific databases, created by K-Dense, that turns Claude Code, Cursor, Codex, Gemini CLI, Google Antigravity, Pi or any host speaking the open Agent Skills standard into something the project calls an AI Scientist. The skills cover biology, chemistry, medicine, physics, engineering, Earth science, data analysis and scientific communication. The repository has 49,945 stars and 4,455 forks, was created in October 2025, and ships version 2.72.0. There is even an arXiv paper describing the collection: Scientific Agent Skills: A Library of Procedural Knowledge for Research Agents.
The phrase "procedural knowledge" is doing the heavy lifting there. Each skill is not a prompt. It is a SKILL.md carrying the purpose, the workflow, the version metadata, the scientific conventions and the validation checks for one package, data source or workflow. Your agent already knows how to import RDKit. The skill tells it which conformer generation settings the field trusts, and what to check before believing the output.
What Is Inside: The 177 Skills in Five Buckets
The collection breaks into five groups, and the counts tell you where the weight is:
| Bucket | Size | What you get |
|---|---|---|
| Scientific and financial databases | 100+ | The Database Lookup skill documents 80 sources with public, registered or licensed access: PubChem, ChEMBL, UniProt, COSMIC, ClinicalTrials.gov and more, with endpoint selection, pagination and provenance. Dedicated skills add DepMap, PrimeKG, AlphaGenome and others |
| Optimized Python package skills | 70+ | Version-aware workflows for RDKit, Scanpy, scikit-learn, PyTorch Lightning, PyTDC, BioPython, Qiskit, GeoPandas, pymatgen, molecular dynamics (OpenMM/MDAnalysis) and more |
| Scientific integration skills | 9 | Benchling, DNAnexus, LatchBio, OMERO, Protocols.io, Open Notebook, Ginkgo Cloud Lab, LabArchives, Opentrons |
| Analysis and communication tools | 30+ | Literature review, evidence-traceable writing, peer review, Paperclip (full-text papers with line-pinned citations), Exa Search, macro-free PPTX posters, Mermaid diagrams |
| Research and clinical tools | 10+ | Hypothesis generation, grant writing, PK/PD modelling and simulation, BIDS, ISO standards-readiness preparation, analytical method validation |
Two things are worth noting about that table. First, the skills are an explicitly defined floor, not a ceiling: the agent can still use any Python package or API, these are just the pre-documented paths that make common workflows faster and more dependable. Second, skills with bundled scripts ship with a test suite under tests/<skill-name>/ and a dependency entry in a shared requirements file, so the collection is maintained like software, not like a folder of prompts.
Why Procedural Knowledge Beats a Clever Prompt
The difference shows up in the small print. A generic agent asked to convert genomic coordinates will happily produce an answer. The Genomic Coordinates skill exists because BED, GFF, GTF, VCF, SAM and WIG files disagree about 0-based versus 1-based indexing and contig naming, and getting that wrong silently corrupts the analysis. The skill catches the mismatch before it happens.
That pattern repeats across the library. The primer design skill runs bounded local or BLAST-assisted specificity screening instead of just returning Primer3 candidates. The QIIME 2 amplicon skill checks read retention through denoising instead of assuming the pipeline worked. The marine carbonate chemistry skill carries measurement uncertainty instead of reporting a single number. Each of these is a mistake a smart non-specialist makes exactly once, and the skill makes sure the agent does not make it at all.
This is also why the collection documents versions. The skills pin baselines like PathML 3.0.8, GeoPandas 1.2.0 and build123d 0.13.0, and each skill records its own compatibility requirements and validation scope. Scientific software moves fast and breaks workflows quietly; version-aware guidance is the difference between a result you can trust and one you have to redo.
Install It in One Command
The standard route installs into every supported host at once:
npx skills add K-Dense-AI/scientific-agent-skills
That covers current versions of Claude Code, Claude Cowork, Codex, Gemini CLI, Google Antigravity and Cursor. If you use the GitHub CLI (v2.90.0 or newer), gh skill gives you more control:
gh skill install K-Dense-AI/scientific-agent-skills
gh skill install K-Dense-AI/scientific-agent-skills scanpy # one skill only
gh skill install K-Dense-AI/scientific-agent-skills --pin v2.71.0 # reproducible
Pinning to a release tag or commit SHA is worth doing for any published workflow, since skill guidance and package baselines both move. gh skill update --all pulls everything current again.
For plugin clients, the repo is also a valid Agent Plugins 1.0.0 package (plugin.json plus skills/), so Cursor, Codex, GitHub Copilot, VS Code, Kiro and other compatible clients can load the whole collection as one plugin. Manual installs follow the ~/.agents/skills/ convention for user-level or .agents/skills/ for project-level. Installing a topical subset keeps the catalog focused: a genomics lab does not need the battery cycling skill in context.
What You Can Actually Ask It to Do
Once installed, the agent picks up relevant skills on its own, or you can name one directly. The README's worked examples read like real lab notebooks:
Drug discovery pipeline. "Query ChEMBL for EGFR inhibitors (IC50 < 50nM), analyze structure-activity relationships with RDKit, generate improved analogs with datamol, perform virtual screening with DiffDock against the AlphaFold EGFR structure, search PubMed for resistance mechanisms, check COSMIC for mutations, and create visualizations and a comprehensive report." One prompt, six skills, a full preclinical prioritization.
Single-cell RNA-seq. The Scanpy skill handles the analysis, Bulk RNA-seq prepares validated counts for a PyDESeq2 handoff, and Pathway Enrichment covers the downstream ORA and GSEA. The handoff boundaries are explicit, which matters: each step's output is validated before the next step consumes it.
Virtual screening campaign, primer design, variant evidence review, systems biology network analysis. The pattern is the same every time: the skill supplies the conventions, the agent supplies the labor, and the validation checks decide whether the result stands.
The Database Layer: 100+ Sources With Provenance
The database coverage deserves its own section because it solves the most tedious part of computational research: finding the right source, hitting the right endpoint, and knowing whether the data can be trusted.
The Database Lookup skill documents 80 databases with their access requirements (public, registered or licensed), endpoint selection, pagination contracts and provenance checks. Dedicated skills go deeper on high-value sources: DepMap for cancer dependency data, PrimeKG for precision medicine knowledge graphs, AlphaGenome for DeepMind's precomputed variant effects across GRCh38. Multi-database packages like BioServices, Biopython and gget extend the reach further.
Provenance is the key word. A number without a source is a rumor, and the skills treat it that way: retrieval contracts say what was fetched, from where, and under what access terms. For anyone building on published analyses, that is the difference between reproducible and decorative.
The Honest Boundaries: What It Refuses to Decide
This is the section that most raises confidence in the project. The skills that touch sensitive territory carry explicit research-only boundaries, stated in the skill, not buried in a disclaimer page:
- Healthcare AI and biosignal skills cover EHR and model research and retrospective validation, explicitly not patient-specific diagnosis, treatment, alarms or deployment decisions.
- Preclinical animal-welfare skills do severity scoring and humane-endpoint forecasting as an aid to assessment, never as a decision rule, with EU Directive 2010/63/EU reporting in mind.
- Regulatory skills prepare draft evidence artifacts for ISO and ICH/USP/CLSI frameworks for qualified review, never as certification, accreditation or method-release decisions.
- Variant effect predictions from AlphaGenome are marked research only, not clinical.
The security disclaimer is equally direct: skills are documentation and helpers, and anything that touches credentials, lab hardware or licensed data sources needs the operator's own review. An agent that knows where its authority ends is more useful than one that never asks.
K-Dense BYOK: The Free AI Co-Scientist on Your Desktop
If you want the skills without assembling the agent setup yourself, K-Dense ships K-Dense BYOK, a free, open-source AI co-scientist that runs on your own desktop. You bring your own API keys, pick from 40+ models, and get a full research workspace with web search, file handling, 100+ scientific databases and all 177 skills from this repo. Your data stays on your computer, and you can optionally scale heavy workloads to cloud compute via Modal.
It is the same playbook as the skills themselves: the labor is automated, the judgment stays with the researcher, and nothing phones home with your data.
Repo Health: 49,945 Stars and an arXiv Paper
The repository was created in October 2025 and had 49,945 stars, 4,455 forks and 32 open issues at the time of writing, with the latest push landing on 12 October 2026. Version 2.72.0 is current. The README is backed by a large docs tree, the blog carries skill benchmarks and deep dives, and CI runs security scans and skill tests on every change.
The durable signals: every skill carries version metadata and a validation scope, breaking changes are documented in release notes, and the project is unusually willing to state what each skill will not do. The commercial side is K-Dense Inc., which offers enterprise support; the skills themselves are MIT licensed with no paid tier.
Good fit if you do computational research in any of the covered domains and you already use an AI coding agent. It is strongest where the conventions are sharp and the mistakes are silent: genomics coordinates, chemistry protonation states, phylogenetics pipelines, PK/PD modelling.
Poor fit if you want the agent to make the call a specialist should make. The skills prepare evidence and draft artifacts; the treatment decision, the certification and the diagnosis stay human, by design. It is also not a substitute for learning the domain: the skills encode the conventions, but you still need to know which question to ask.
FAQ
What is Scientific Agent Skills?
Scientific Agent Skills is a free, MIT-licensed library of 177 validated skills plus 100+ scientific databases for AI agents, created by K-Dense. Each skill gives an agent the procedural knowledge, conventions and validation checks for one scientific package, data source or workflow, covering biology, chemistry, medicine, physics, engineering, Earth science and scientific communication.
Which AI agents does it work with?
Any host that supports the open Agent Skills standard: Claude Code, Cursor, Codex, Gemini CLI, Google Antigravity, Pi, OpenClaw and others. It is also packaged as an Agent Plugins 1.0.0 plugin for Cursor, Codex, GitHub Copilot, VS Code and Kiro. Install with npx skills add K-Dense-AI/scientific-agent-skills.
Is Scientific Agent Skills really free?
Yes. The collection is MIT licensed with no paid tier. K-Dense offers commercial enterprise support separately, and the optional K-Dense BYOK desktop app is also free and open source; you only bring your own model API keys.
Can it diagnose patients or certify lab methods?
No, and the skills say so explicitly. Healthcare skills are for research and retrospective validation, not diagnosis or treatment decisions. Regulatory skills prepare draft artifacts for qualified human review, never certification or accreditation. The judgment stays with the researcher.
How is this different from just asking my AI agent to do science?
A generic agent can call any library, but it does not know the field's conventions: 0-based versus 1-based genomic coordinates, validation handoffs between analysis steps, or which package versions a workflow was tested against. The skills encode those conventions with version-aware guidance and validation checks, so the agent's output can be trusted, not just generated.
How do I install only the skills I need?
With the GitHub CLI: gh skill install K-Dense-AI/scientific-agent-skills scanpy installs one skill. Or clone the repo and copy selected skill folders into ~/.agents/skills/. Pinning with --pin v2.71.0 keeps published workflows reproducible.
Originally published on DevPik.
Top comments (0)