Scientific Agent Skills scientific AI agent: a guide

Scientific Agent Skills scientific AI agent: 177 MIT-licensed skills, install steps from the README, the project’s security report and usage limits.
Scientific Agent Skills agent IA scientifique : guide

Scientific Agent Skills scientific AI agent: the K-Dense-AI/scientific-agent-skills repository gathers 177 “skills” (folders of instructions with a SKILL.md file) that teach an AI agent to use research libraries, databases and tools. It showed about 53,700 stars and 4,700 forks in October 2026, read on the repository page on 22 October. The project used to be called Claude Scientific Skills; it now presents itself as compatible with any agent that follows the open Agent Skills standard.

For an SRE or platform owner this repository is also a textbook case: 177 packages of instructions and scripts that an agent can run on a lab machine. You need to know what you install, with which rights, and what the project itself publishes about its security.

The information comes from the README and the repository files, read on 22 October 2026. span.observer did not install these skills: the images below reproduce real excerpts from the repository.

Scientific Agent Skills scientific AI agent: what the repository says

  • Who: K-Dense (K-Dense Inc., the K-Dense-AI organisation on GitHub). The LICENSE file reads “Copyright (c) 2025 K-Dense Inc.”.
  • Licence: MIT for the repository. The README states that each skill may carry its own licence, declared in its SKILL.md: check them one by one.
  • Version: 2.72.0 according to the README. The repository page shows no release date.
  • Content: 177 skills; the README claims more than 100 scientific databases, 80 of them documented in the “Database Lookup” skill, and about twenty categories (bioinformatics and genomics, chemistry, medical imaging, data, scientific communication and so on). The README also cites a research paper (arXiv:2609.00065) describing the collection.
  • Compatible agents: Cursor, Claude Code, Codex, Gemini CLI, Google Antigravity, GitHub Copilot, VS Code, Kiro, OpenClaw, Pi, Hermes. Install paths vary by host.
  • Recent change: version 2.72.0 no longer bundles the docx, pdf, pptx and xlsx skills, which you must install from Anthropic’s repository if you need them.

Anatomy of a skill

Each skill is a folder under skills/ with a SKILL.md: a YAML header (name, description, licence, version) followed by Markdown instructions. Some add references/ and scripts/ folders.

Start of the SKILL.md file of the database-lookup skill in the Scientific Agent Skills repository: YAML header with name, description, allowed tools, MIT licence and version 1.8, then the six workflow steps

The “Database Lookup” skill sets the tone: it tells the agent to define a retrieval contract, to check counts when completeness matters, and to treat API responses as untrusted data. These are good agent practices, but they are instructions, not guarantees.

Installing Scientific Agent Skills from the README

Step 1 — Choose a method

The README offers three options. The shortest uses npx, the second uses the GitHub CLI (version 2.90.0 or later):

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

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

The second command installs a single skill (here scanpy). The README also describes --agent to target a specific host and gh skill update --all for updates.

Step 2 — Pin the version

gh skill install K-Dense-AI/scientific-agent-skills --pin v2.71.0

The README advises pinning a tag or a commit for reproducible installs (its example uses v2.71.0).

Step 3 — Scan third-party skills before installing

uv pip install cisco-ai-skill-scanner
skill-scanner scan /path/to/skill --use-behavioral
Excerpt from the Scientific Agent Skills README: npx skills add, gh skill install with a single skill and with --pin, then installing the cisco-ai-skill-scanner and the skill-scanner scan command

The third option (Cursor or Codex plugin) links or installs the repository as an “Agent Plugins” package; see the README for the exact commands. The README notes that installing a skill’s files does not install its dependencies: plan one environment per skill.

Security: read the project’s own report

A skill can run code, install packages, make network requests and change files: the README says so and asks you to read each SKILL.md, install only what you need and scan third-party skills. The project scans its skills weekly with Cisco AI Defense Skill Scanner and publishes the result in docs/security-report.md.

Excerpt from docs/security-report.md in the Scientific Agent Skills repository: report generated on 19 October 2026, 177 skills scanned, 466 findings including 32 critical and 21 high, then a table of six skills with their severity

The 19 October 2026 report lists 177 skills scanned, 466 findings, including 32 critical and 21 high, and 159 of 177 skills rated safe. There are two ways to read this. Publishing these numbers is a sign of transparency. On the other hand, the README itself warns that a clean result does not guarantee the absence of risk, and that contributed skills are also reviewed manually before merging: read the findings for the skills you plan to use rather than trusting the total.

How to observe it in use

The repository is a library, not a service: it emits no traces or metrics. What you monitor is the agent that uses it. A few pointers:

  • Traces and costs. Measure the agent’s model and tool calls as in AI agent cost with OpenTelemetry: a skill that triggers long searches in public databases can push consumption up.
  • Rights. Run the agent in a restricted environment; the README notes that in NemoClaw, skills that need the network only work after the operator approves the domains. For isolation, see NVIDIA OpenShell and guardrails in production.
  • Provenance. Keep the list of installed skills with their version (metadata.version) and the last-review date that SKILL.md shows.

Limits and open points

  • Research use only. Several clinical or regulatory skills are explicitly research-only: the README rules out diagnosis, treatment, patient-specific decisions and medical-device validation. Qualified professionals remain responsible.
  • Multiple licences. MIT for the repository, but each skill may carry another; check before any commercial use.
  • Project figures. The number of skills and databases is a README claim; it changes between versions (181 skills in 2.71.0, 177 in 2.72.0).
  • Dependencies. Each skill may need its own packages, versions or API keys.

Best practices

  • Install only the skills you need, after reading their SKILL.md.
  • Pin a version and read the changelog before updating.
  • Scan third-party skills and read the project’s security report for those you keep.
  • Separate API keys per skill and per environment.
  • Never present an agent’s output as a scientific result without human verification.

Frequently asked questions

Do I need Claude Code to use these skills?

No. According to the README, the collection follows the Agent Skills standard and works with several agents (Cursor, Codex, Gemini CLI and others). For Claude Code’s plugin directory, see claude-plugins-official.

Is the project free?

The code is MIT-licensed; the models, APIs and databases the agent queries have their own terms and may be paid.

Can it be used in a clinical setting?

The README says several skills are research-only and do not support diagnosis or treatment. Do not use them for decisions about patients.

Conclusion

Scientific Agent Skills is a large library of instructions for research agents, published with a regular security report, which is rather rare. It is still code and instructions that run with your rights: pin, scan, restrict, and keep a qualified human to interpret results. For another example of security-oriented skills, see the Cloudflare security audit skill.

span.observer is an independent community blog, not affiliated with Dynatrace LLC or K-Dense. Trademarks belong to their owners. Tools change quickly: check commands in the official README (github.com/K-Dense-AI/scientific-agent-skills).

Total
0
Shares
Leave a Reply

Your email address will not be published. Required fields are marked *

Previous Post
Claude plugins official répertoire plugins : guide

Claude plugins official plugin directory: a guide

Next Post
Ruflo essaims agents IA : installer et superviser

Ruflo AI agent swarms: install and monitor

Related Posts