SkillLib

21,516 leading agent skills, ready to search.

SkillLib collects public repositories of Anthropic-compatible (Claude) and OpenAI-compatible (Codex) skills, keeps one skill to follow for each overlapping name, and lets an agent find and load a procedure without putting the library in the prompt.

MIT · Python 3.11+ · no third-party dependencies

Project    Homepage    PyPI

Install

Python 3.11 or newer. SkillLib has no runtime dependencies.

Download the package from pypi.org/project/skilllib.

python -m pip install skilllib

That install provides the library and the skilllib command. The skill catalog is the catalog/ directory in the git repository. Commands and the Python functions read that directory from the current working directory, so a search runs from a checkout:

git clone https://github.com/DistriAI/SkillLib.git
cd SkillLib
skilllib find "Run GSEA on a ranked gene list"

To work on the library itself, install the checkout in place with python -m pip install -e .. To use the checkout without installing it:

export PYTHONPATH=src
python -m skilllib find "Run GSEA on a ranked gene list"

search, add, and update call GitHub. visit checks a repository out with git. Judging a new overlap with a model uses the OpenAI-compatible endpoint in catalog/judge_config.json (SKILLLIB_LLM_BASE_URL, SKILLLIB_LLM_MODEL, and SKILLLIB_LLM_API_KEY override it). update --judge heuristic judges from the repository metadata alone.

The library

The catalog is the public skill-content repositories with more than 1,000 GitHub stars, indexed on 16 October 2026. A match is a repository full of task skills (SKILL.md), or a curated collection that points at those repositories. Skill specifications, skill SDKs, and agent products are left out.

32,977	skill files in the source repositories
21,516	leading skills an agent can be offered
179	skill repositories, and 15 collections
13	domains, 13,320 scenarios
389	task families, 964 product channels
8,902	duplicate copies kept as variants, not as peers

A separate skill is the only copy of that name. An authentic skill is the copy to follow when several copies agree. A combined skill is what remains when the copies disagree: the conflict note says how to proceed, and each peer keeps its repository and path.

Use it

The integration is three calls.

from skilllib import find_skills, load_skill

found = find_skills("Run GSEA on a ranked gene list")
skill = load_skill(found["cards"][0]["id"])

skilllib find "Run GSEA on a ranked gene list"
skilllib load combined::gsea
skilllib prompt --working-set catalog/working_set.example.json

find_skills returns a few cards for the task: name, description, when to use the skill, when not to, and where it came from. load_skill returns the procedure. render_prompt writes the standing prompt for those Python function names: one sentence for each of the 13 domains, an instruction to search before deciding that no skill applies, and a cap of three loaded bodies.

Integrating an agent

An agent runtime takes the prompt and the tool descriptions from skilllib.integrate: tool_specs(), resident_prompt(), task_instruction(), and call_tool(). Register the specs under the names they already carry (skill_library_find, skill_library_load, skill_library_visit), append resident_prompt() to the system prompt, and return call_tool(name, arguments, root=checkout) to the model. The field reference is docs/interface.md.

An agent that already has skill_view (Hermes) or the equivalent Claude or Codex skill loader should keep it. That tool opens a skill installed on the machine, including its scripts and references. SkillLib answers the prior question, at a size those lists are not built for: which skill from the library is worth installing. load_skill can show the cached text before that install. When the skill names bundled files that are not checked out, the tool result names the repository to visit instead of pretending the procedure can run.

On a hand-labeled set of 40 tasks, the finder returned the expected skill for 36.

Shelves

Pass a domain to search one shelf: find_skills(task, domain="science").

Domain	Scenarios	Skills
research	1,527	4,997
science	1,609	2,446
software-engineering	1,690	2,146
business	1,539	2,134
security	1,418	1,787
agent-tooling	882	1,478
marketing	957	1,347
devops	928	1,106
general	638	1,020
data	683	889
design	594	872
media	423	671
documents	432	623

Regenerate the local shelf with python -m skilllib browse. Open html/index.html. The web homepage is distriai.github.io/SkillLib.

Grow the catalog

New repositories merge in. A later search does not throw away the 1,000-star collection.

python -m skilllib add owner/name
python -m skilllib update
python -m skilllib search --min-stars 500 --merge

update judges a name again only when a copy joins it, leaves it, or a stored skill file changes. Rebuild the task families when you want the shelves collapsed again. That is a separate command, because re-collapsing every domain on each added repository is expensive:

python -m skilllib families --name-with heuristic

Commands

Run these from the repository root as skilllib <command>, or as python -m skilllib <command>. Full arguments are in docs/interface.md. The stage log for the first catalog is in docs/history.md.

Command	What it does
search	Find skill repositories on GitHub. --merge keeps the ones already stored.
add owner/name	Inspect one repository and merge it into the catalog.
list	Print the saved catalog.
update	Index skills added or edited since the last update.
find "..."	Return a few routing cards for a task.
load <id>	Print the cached procedure for one card.
prompt	Write the standing prompt fragment.
families	Collapse the current cards into families again.
browse	Write the HTML shelf under html/.
visit owner/name	Check out one repository as a submodule.

screen, download, index, cards, enrich, and link run a single stage. download --leaders fetches the skill file behind every leading skill, including a name that occurs once. update runs the stages that a new repository needs.

License

This project is MIT licensed. Skill files collected from other repositories keep the license recorded for each repository in the catalog.