Claude Skills Library is a third-party directory that indexes agent skills published on GitHub and makes them browsable by task, with a copyable install command and a ZIP download on every listing. Browsing is free and needs no account.
Used this tool? Rate it
Used this tool? Rate it
Claude Skills Library is a third-party web directory that indexes agent skills published on GitHub and makes them searchable by task. It is not operated by the company that publishes the agent platform those skills run on. The operator describes itself as a small team of developers, and states that listings are sourced from the community through GitHub and reviewed before they go live. Understanding that boundary first saves confusion later: the catalogue curates and links, but the skills themselves belong to other people and live in other repositories.
A skill, in the sense this catalogue uses the word, is a modular instruction set an agent can load to carry out a specific task on its own. That definition matters because it separates a skill from a saved prompt. A skill ships as a file with structured metadata, often alongside scripts and reference material, and an agent decides when to pull it in. The directory's job is discovery: it reads those files from public repositories, groups them by what they do, and gives each one a page.
The site presents itself as a place to browse by use case, download in minutes, or contribute your own through GitHub. Those three verbs describe the whole product. There is no runtime here, no hosted execution, and no account system layered on top. What you get is a search surface over a large body of public files, plus the plumbing to move one of those files onto your machine quickly.
Seventy-two category pages split the catalogue by task, and the per-category counters vary by two orders of magnitude. Code generation, architecture design and testing sit in the thousands; lead research, ticket triage and response drafting sit in the tens, and the smallest categories such as escalations and handoffs hold only a handful. That spread is the most useful signal the browse interface gives you, because it tells you in advance whether a category is a real corpus worth filtering or a thin shelf you can read end to end in a few minutes.
Alongside categories, the site runs a leaderboard and a featured shelf. The leaderboard ranks skills by community installs and names the source repository next to each entry, which lets you see at a glance how concentrated the ecosystem is: a handful of publisher repositories account for a large share of the top ranks.
Every listing keeps a visible line back to the GitHub repository it was indexed from, so the directory never becomes the only copy of a skill. Two retrieval routes sit side by side on each page. One is a copyable install command that pulls the skill straight from its source repository. The other is a ZIP download of the SKILL.md file and its bundled resources, meant for use with Claude Code or other agents.
The difference between those two routes is worth a moment. The install command is faster and keeps a link to upstream, which matters when the author ships a fix. The ZIP is a snapshot; it works offline and in environments where you would rather not run an installer, but it will not update itself and it will drift from upstream silently.
Contributions go through a security audit and then a pull request that a maintainer has to merge, rather than publishing straight to the catalogue. A repository has to contain a valid SKILL.md file to qualify at all, and anything that ships executable scripts draws an extra security pass. The site also states a preference for permissive licences such as MIT or Apache 2.0, though this is guidance rather than a hard admission rule.
This puts the directory in the middle of the curation spectrum. It is not a raw crawler that indexes anything shaped like a skill, and it is not a hand-picked shelf of a few dozen editorially chosen entries either. A human sits in the merge path, but at catalogue scale that human is checking admission, not auditing every instruction.
The site exposes a public JSON endpoint over the same catalogue the browse pages use, returning per-skill records with slug, description, source repository, star and fork counts, a score, and a pagination object. Nothing on the site advertises this as a supported API, so treat it as an implementation detail that can change without notice rather than an interface to build on.
The practical path from a search box to a working skill is short, and it is worth walking through once. The four patterns below are the ones the catalogue's own structure supports directly.
A skill is instructions your agent will follow, which makes review a different exercise from reading a library's README. Admission already filters some of this: a valid SKILL.md file is required to qualify at all, and submissions that ship scripts get an additional security review before a maintainer merges them. Passing that filter is not the same as being right for your project. Two checks carry most of the weight. First, does the skill bundle executable scripts? Those get a stated extra security pass on submission, but they are also the part that runs on your machine. Second, does the instruction body ask the agent to fetch remote content or reach for credentials? Neither is automatically wrong, but both change what a mistake costs you.
The people who get the most out of this catalogue are the ones who can read a SKILL.md file and judge it themselves. That is the load-bearing requirement, and it follows directly from how the service is structured. Community sourcing through GitHub plus a pre-publication review by a small maintainer team is a filter, not a warranty, so the reading stays with you. Scale shapes the fit too: eighty-four category pages split the catalogue by task, and the per-category counters vary by two orders of magnitude, which means your own field may be a deep corpus or a nearly empty shelf.
It fits less well in three situations. Organisations that need vetted, contractually supported components will not find that here, because the terms disclaim exactly those guarantees. Teams that need private skills, role-based access or an audit trail need a registry with those features rather than a public index. And anyone who wants a small, first-party-blessed set will be better served going directly to the platform vendor's own repositories, where the catalogue is tiny by design and reviewed rather than aggregated.
The directory itself is a plain website with no desktop application, no command-line client of its own and no browser extension. Anything with a browser reaches it, and nothing needs to be installed to browse.
Reach on the consumption side is broader than the name suggests. The site names Cursor, Codex and AntiGravity alongside Claude Code as environments its listings work in. That breadth is a property of the file format rather than of the directory: a SKILL.md file is markdown plus metadata, so any agent that reads the convention can use it.
The install command depends on Node.js being available, since it runs through npx. The ZIP route has no such dependency, which is the practical reason to keep both paths in mind rather than defaulting to one.
| Route | What you need | Stays linked to upstream | Works offline |
|---|---|---|---|
| Install command | Node.js and network access | Yes | No |
| ZIP download | A browser only | No | Yes |
| Manual clone from source | Git and network access | Yes | No |
Browsing, searching and downloading require no account, and the catalogue carries no paid tier, no per-skill charge and no subscription. The homepage FAQ states most skills in its database are free and open source. Revenue instead comes from advertising: the privacy policy discloses that the site serves Google AdSense, which may set its own cookies.
Cost, then, is not the axis on which you choose here. The real cost is review time, and it scales with how many skills you adopt rather than with any billing meter.
The realistic alternatives fall into a few shapes, and the shape matters more than the headline count. Independent surveys of this landscape group the options into registries with accounts and versioning, install-count leaderboards, automated aggregators that crawl GitHub wholesale, first-party example repositories, and hand-curated lists. They differ mainly in curation depth and security posture, not in what a skill file looks like.
No third-party catalogue in this space is affiliated with the company that publishes the agent platform itself. That single fact reframes the comparison: the choice is not between an official option and unofficial ones, but between differently curated third-party indexes plus a small first-party repository that covers only what its publisher wrote. If you need versioning, private skills or access control, a registry is the right shape. If you want the smallest reviewed set, go first-party. If you want breadth with a human in the merge path, a curated directory like this one sits between those poles.
The operator states plainly that it does not endorse, warrant or guarantee the quality, security or fitness of any listed skill. This is the single most important line in the terms, and it is not boilerplate: it follows from the structure, since listed skills stay with their authors on third-party platforms. The same section disclaims warranties of uptime, accuracy and suitability, offering the site as is.
The two published catalogue figures do not agree with each other, and the gap is large enough to change what you expect from a search. The homepage FAQ describes a database of more than 90,000 skills, while the site's own public endpoint reports a total of 49,878 published records. Both numbers are stated by the operator; neither has been reconciled publicly. Read the larger figure as marketing copy and the smaller one as the count the running system reports.
The install leaderboard is a genuine adoption signal, but it carries its own stated refresh date and that date can sit months behind the catalogue. Popularity is also not review — a heavily installed skill is evidence that people tried it, not that anyone audited it.
Two further boundaries are worth naming. Corrections and takedown requests run through issues on the operator's public repository rather than an in-product report button, so getting a wrong listing fixed is a public, asynchronous process. And the licence that governs your use of a downloaded skill is the author's licence, not anything the directory grants; the directory owns only its own interface and curation layer.
On privacy, reading the catalogue is anonymous: measurement runs through a cookie-free analytics tool that records page views, referrers and coarse region only, and the operator states it does not collect personally identifiable information from visitors who merely browse. Submitting a skill is different, since the form collects a contact email retained for as long as the listing stays up. Anonymised analytics are kept for up to twelve months. Visitors in the EU, EEA and UK are told they hold access, rectification and erasure rights, exercisable through the contact page.
No. It is an independently operated directory, and independent comparisons of this landscape confirm that no catalogue of this kind is affiliated with the platform vendor. The vendor distributes its own skills through its plugin marketplace and GitHub repositories.
No. Browsing, searching, reading a rendered SKILL.md and downloading are all available without registering. An email address is only collected if you submit a skill.
The site states two different figures. Its homepage FAQ claims a database of more than 90,000; its public endpoint reports 49,878 published records. Treat the smaller number as the one the live system reports.
The homepage states most of them are free and open source. Because each skill stays under its author's own licence, the answer for any single skill is whatever that repository's licence says.
It runs a security audit on submission and applies an extra security review to skills containing scripts, and a maintainer merges the pull request. But the terms explicitly decline to guarantee security or fitness, so the audit is an admission filter rather than an assurance.
The site names Claude Code, Codex, Cursor and AntiGravity. The format is markdown with structured metadata, so compatibility is a property of the agent reading it rather than of the directory.
By opening an issue on the operator's GitHub repository, which is the stated route for bug reports, incorrect skill information and takedown requests.
It is a good way to see what people install, which is not the same as what is well written or well maintained. It also carries a stated refresh date that may lag the catalogue, so use it as one input rather than a ranking to follow.
