Scientific Agent Skills

Scientific Agent Skills is K-Dense's large Agent Skills collection for science, research, engineering, data analysis, and scientific communication. The repository packages SKILL.md instruction files, references, scripts, and examples so compatible agents can perform bounded scientific workflows instead of improvising from generic Python or web-search knowledge.

Use this listing for the skill pack itself. Use separate tool, MCP, or package entries when evaluating a runtime, local AI workspace, database service, scientific Python package, or cloud execution platform.

Knowledge Freshness

Verified on 2026-06-18, the upstream repository reported version 2.52.0 in pyproject.toml, a latest GitHub release v2.52.0 published on 2026-06-12, and recent repository activity on 2026-06-15. The README described 147 ready-to-use skills and 100+ scientific database or data-access paths.

The pack changes often and includes both maintainer-authored and community contributed skills. Before running a workflow, check the exact installed release, read the relevant SKILL.md, and verify current database, package, and API documentation for the specific scientific domain.

Retrieval Sources

This listing is grounded in:

• The upstream K-Dense-AI/scientific-agent-skills README.
• The generated docs/skills.md skills catalog.
• pyproject.toml version and Python tooling metadata.
• The repository SECURITY.md scan summary.
• Representative database-lookup and clinical-decision-support SKILL.md files.
• Current GitHub repository metadata for license, stars, release, and topics.

Core Workflow

Install through the open Agent Skills installer:

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

GitHub CLI users can also use the documented gh skill path:

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

For safer production use, install or enable only the topical subset needed for the job. A good workflow is:

1. Pick the scientific domain and exact task boundary.
2. Read the relevant SKILL.md and references before execution.
3. Confirm required packages, API keys, data licenses, and outbound network access.
4. Run on a copied or versioned workspace first.
5. Review code, retrieval logs, citations, figures, tables, and conclusions before using the output.

Capability Scope

Use Cases

• Give Claude Code, Codex, Cursor, OpenClaw, or another compatible agent a source-backed research workflow for a specific scientific domain.
• Query public scientific databases with explicit endpoints, pagination, identifier conversions, and provenance instead of unbounded web summaries.
• Build reproducible analysis scaffolding for Scanpy, RDKit, BioPython, scikit-learn, Dask, molecular dynamics, statistical analysis, or related Python workflows.
• Draft scientific manuscripts, literature reviews, posters, slide decks, or peer-review notes with citations and reporting-guideline awareness.
• Create controlled internal workflows for clinical research, cohort analysis, clinical-trial evidence synthesis, drug discovery, or laboratory automation.

Production Rules

• Prefer topical skill installs over installing the full repository.
• Treat all generated code, data transformations, citations, clinical conclusions, and scientific claims as drafts until reviewed by a qualified human.
• Read the specific SKILL.md and reference files for the skill being used; do not rely on the repository README as the only operating guide.
• Validate identifiers, organism names, genome builds, chemical structures, clinical-trial filters, units, statistical assumptions, and database access dates before accepting results.
• Do not run skills against PHI, controlled datasets, unpublished lab results, proprietary assays, credentials, or production lab systems unless the workspace, model provider, and network path are approved for that data.
• Review upstream security-scan findings before enabling skills that run scripts, call external LLMs, access environment variables, capture screen context, write files, or interact with lab and clinical systems.

Safety Review

The upstream project is unusually valuable because it is broad and current, but that breadth is also the main risk. The 2026-06-15 upstream security summary reported critical and high findings in multiple skills. Some findings relate to environment-variable access, network calls, generated skill content, screen-content capture, prompt-injection exposure, and clinical or research document generation.

That does not make the whole repository unusable, but it means operators should treat it like a high-power research automation pack: install narrowly, read before running, keep agent permissions scoped, and require human review before any scientific, clinical, regulatory, laboratory, or publication-facing output is trusted.

Source Review

Verified on 2026-06-18:

• GitHub metadata reported K-Dense-AI/scientific-agent-skills as an MIT-licensed repository with topics including ai-scientist, bioinformatics, chemoinformatics, clinical-research, drug-discovery, genomics, proteomics, scientific-computing, and agent-skills.
• The README described Scientific Agent Skills as an Agent Skills collection for Cursor, Claude Code, Codex, Google Antigravity, OpenClaw, Pi, Hermes, and other compatible agents.
• pyproject.toml declared package version 2.52.0, Python >=3.13, and dependencies including cisco-ai-skill-scanner, firecrawl-py, and python-dotenv.
• docs/skills.md described coverage across scientific databases, lab and R&D integrations, Python scientific packages, research methodology, scientific writing, document processing, laboratory automation, and computational-resource discovery.
• skills/database-lookup/SKILL.md described deterministic querying across public scientific, biomedical, materials, regulatory, finance, demographic, astronomy, environmental, chemistry, drug, protein, gene, clinical-trial, patent, and economic databases.
• skills/clinical-decision-support/SKILL.md described group-level clinical research and evidence-synthesis documents, while explicitly distinguishing those documents from individual bedside treatment plans.
• SECURITY.md reported 147 scanned skills, 900 total findings, 67 critical findings, 43 high findings, and 107 of 147 skills marked safe in the generated 2026-06-15 scan summary.