scientific-agent-skills: 177 Skills for AI Agents

TL;DR: scientific-agent-skills is a Python library that packages 177 pre-written agent skills and more than 100 scientific databases into an AI coding agent. It is MIT licensed with 49,407 stars, 4,424 forks and only 31 open issues.

What it is

Agent Skills are a packaging format: instead of a system prompt you hand an agent, you hand it a directory of procedures it can pull in when a task calls for them. This repository is a large library of those, aimed squarely at scientific work.

The repository description makes three concrete claims, and they are the project’s own words rather than anything independently verified: the number one Agent Skills library for science, usage by 250,000 or more scientists, and 177 ready-to-use validated skills plus over 100 scientific databases covering biology, chemistry, medicine and drug discovery. The domains listed are the useful part. “Science” as a category is vague; a bioinformatics skill and a chemoinformatics skill are concrete workloads you can point an agent at.

$ curl -s https://api.github.com/repos/K-Dense-AI/scientific-agent-skills \
  | jq '{stars: .stargazers_count, language, license: .license.spdx_id, topics, created_at}'
{
  "stars": 49407,
  "language": "Python",
  "license": "MIT",
  "topics": ["agent-skills","ai-scientist","bioinformatics","chemoinformatics","claude","claude-skills"],
  "created_at": "2025-10-19T20:54:15Z"
}

The compatibility claim is the interesting part

The description names Cursor, Claude Code, Codex, Pi, Antigravity, and the open Agent Skills standard. That is a much broader claim than “works with one client”, and if it holds it means the library is not locked to a single vendor’s directory layout.

Read the topic tags against that claim and you get a fair amount of tension. Every Claude-flavoured topic is there, alongside the generic agent-skills tag. The format was clearly shaped inside the Claude ecosystem, and the cross-client support is the claim to test rather than the claim to inherit. Each client listed is worth a five minute check before you plan around it.

The issue count is striking

31 open issues against 49,407 stars is about 0.06 percent, and there are 4,424 forks. For a project created on 2025-10-19 that is either unusually fast issue handling or a preference for handling discussion somewhere other than the tracker. Either way, do not read the low count as a signal that the library is finished.

Metric	Value
Stars	49,407
Forks	4,424
Open issues	31
Primary language	Python
Created	2025-10-19
Last push	2026-10-12
License	MIT

The homepage points at an arXiv preprint rather than a product site. For a project making quantitative claims about its own reach, that is the document to read first.

How to try it

git clone https://github.com/K-Dense-AI/scientific-agent-skills.git
cd scientific-agent-skills

Start by confirming that your agent discovers the skills directory in the layout it expects, then pick one skill from your own field and run it against work you already know the answer to. A skills library is only useful if its procedures survive contact with your actual data.

When to use it

It earns its place when your agent work is domain-heavy and the domain knowledge is what is missing. Drug discovery pipelines, bioinformatics wrangling, chemistry workflows: these are cases where an agent knows how to write code but not what a valid analysis looks like, and that gap is exactly what a curated skill set closes.

It is also a reasonable template if you maintain your own agent skills, since a library at this size shows how the format is meant to be organised.

When not to use it

If your client does not implement the Agent Skills format, none of this applies and you are reading the documentation for a tool you cannot run.

Be careful with the word “validated” in the description. It is a claim by the repository about its own contents, and no benchmark or audit is cited in the metadata. If a workflow feeds results into something regulated, run the output past your own review before trusting the label.

Source and Accuracy Notes

• Project page: https://arxiv.org/abs/2609.00065
• Source repository: K-Dense-AI/scientific-agent-skills (verified via GitHub API html_url)
• License: MIT (verified via GitHub API license.spdx_id)
• Stars at time of writing: 49407
• Skill count, database count and user count are the repository’s own claims from its description, not independently verified