What is an agent skills directory?

A site that indexes SKILL.md files so you can find and install them. Directories differ in where they get skills from (crawling GitHub, editorial selection, or author submissions), what they check before listing one, and how you install from them.

Which agent skills directory is the largest?

Nobody can answer this honestly. Every directory self-reports its count, the numbers are mutually inconsistent, and several have published four different figures in the same year. Treat size claims, including ours, as marketing rather than measurement.

Do these directories host the same skills?

Largely yes. Most index public GitHub repositories, so the same popular skill appears in many of them. Duplication across a single directory is common too: roughly half of the SKILL.md files on GitHub are copies of another file.

Is it safe to install a skill from a directory?

A skill is instructions your agent will follow, and some carry scripts. Check what a skill can do before installing it, prefer directories that publish what they checked and how, and read the SKILL.md itself for anything from an author you do not know.

The agent skills directory landscape in 2026

There are now more than twenty public directories of Agent Skills. They index largely the same GitHub repositories, publish mutually contradictory counts, and describe themselves in nearly identical language. This is an attempt at an honest map: what each one is, what it is genuinely good at, and where it falls short.

We run one of them. That is a conflict of interest, so the ground rules are stated up front and the table records where we lose.

Ground rules

Counts are claims, not measurements. Every figure below is what the site says about itself. They do not reconcile. skills.sh alone has been cited at 57,000, 87,000, 600,000 and 1.37 million within 2026. A directory that indexes GitHub can inflate its number by counting duplicates, forks, and files that are not really skills, and none of them publish a method. Our own figure is in the same category, so read the column as “what they say” and not as a ranking.

Duplication is the reason the numbers are meaningless. Measured across a full crawl of public GitHub, about 2.09 million files named SKILL.md exist but only 1.098 million are unique. Roughly 47 percent are byte-identical copies of another file. Any directory that has not deduplicated is reporting the larger number.

Inclusion criteria. Public, reachable in September 2026, and indexing Agent Skills rather than MCP servers or prompts. Private and enterprise registries are covered in a separate section because they are a different product.

The directories

Directory	Claims	What it is	Review published	Install path
skillsdirectory.com	573,245	”Security first” positioning, security grades and author badges	Yes: 120 detection patterns across 10 threat categories, method not published	Copy, submit path for authors
skills.sh	1.37M	Vercel Labs. Install-count leaderboard and 8-week trend data across 18 agent platforms	No	`skills` CLI
skillsmp.com	up to 2M	GitHub aggregator organised by 867 occupations	No	Copy
skillmd.com	860,000+	Open registry. Format linting, capability flags, third-party scanner results published per skill	Yes: own lint plus NVIDIA SkillSpector and Cisco AI Defense Skill Scanner, results public	CLI, MCP server, Claude Code plugin marketplace, GitHub Action
lobehub.com/skills	169k to 230k	Polished marketplace inside a larger product	No	Copy
agentskill.sh	69k to 110k	Cross-platform index covering 20+ tools	No	Copy
clawhub.ai	20,000+	OpenClaw’s official registry, vector search	Partial	CLI
explainx.ai/skills	10,000+	“Curated, ranked, vetted” with a public submit path	Claimed, not published	Copy
skillsplayground.com	8,884	Open directory, also indexes MCP servers	No	Copy
agenticskills.io	193	Editorial selection rather than crawling. Also publishes a directory of directories	Editorial	Copy
agentskills.codes	not published	Category browsing and a popularity leaderboard	No	Copy
skillregistry.io	not published	Positions itself as the definitive SKILL.md registry	No	Copy
claudemarketplaces.com	2,500+ marketplaces	Indexes Claude Code plugin marketplaces rather than individual skills	No	`/plugin marketplace add`
localskills.sh	n/a	Private skills, versioning, team roles. Sells the angle a public directory cannot	n/a	CLI

Figures as stated by each site in September 2026.

What the table does not show

Almost nobody publishes a review method. Two directories claim security review. One publishes pattern counts without the patterns. We publish third-party scanner output per skill at /auditors, which is checkable, but our coverage is partial and still being backfilled, and that is a real limitation rather than a footnote.

Licence data is missing almost everywhere. Across the skills we index, only 16.6 percent declare a licence. That means 83.4 percent of Agent Skills on the internet tell you nothing about whether you may use them. No directory surfaces this, ours included until recently. If you are installing skills into a commercial codebase, this is the finding that should worry you, not the size of anyone’s index. There is more on this in how to evaluate a skills registry.

Install is where the directories actually differ. Most of them give you a copy button. That is fine for one skill and miserable for twelve. The ones with a real install path are skills.sh, clawhub, ours, and anything speaking the Claude Code plugin format.

Where each one is the right answer

skills.sh if you want to know what people actually use. The install-count leaderboard and trend data are genuinely useful signal and nobody else has them. It is backed by Vercel Labs, so the CLI is well built.

skillsdirectory.com if security posture is your first filter and you are willing to trust an unpublished method. It ranks first for the obvious queries and has the cleanest presentation of grades.

agenticskills.io if you would rather read 193 skills someone chose than search 800,000. Editorial selection is underrated when you do not know what you are looking for.

clawhub.ai if you are in the OpenClaw ecosystem. It is the official registry and the vector search is good.

localskills.sh if your skills are private. None of the public directories solve team-internal distribution, and this one is built for it.

lobehub.com/skills if you already use LobeHub. The integration is the point.

skillmd.com, ours, if you want the SKILL.md itself checked before you install it, if you need something other than a copy button (CLI, MCP, plugin marketplace, CI), or if you care about provenance and licence. We lose on curation against agenticskills.io, on usage signal against skills.sh, and on presentation against LobeHub.

The layer above all of this

Agent Skills are an open standard published at agentskills.io and stewarded through the Agentic AI Foundation, a Linux Foundation body co-founded by Anthropic, OpenAI and Block with more than 170 member organisations. The specification repository has over 25,000 stars and roughly 47 conformant products.

No directory owns the format. Any of us can be replaced, and a SKILL.md you install today keeps working if we are. That is the strongest argument for the standard and the weakest argument for brand loyalty to any index, ours included.

How to choose

Ask three questions in this order. Does it have the skill I need? Does it tell me anything about that skill that I could not learn by opening the file myself? Can I install it without pasting twelve commands?

Most directories answer yes, no, no. The first question is nearly always satisfied because they all index the same repositories. The second and third are where the real difference is, and they are the subject of how to evaluate a skills registry.

If you are specifically on Claude Code, where to find skills for Claude Code covers the install path for each directory. If you are deciding between a public index, a curated list and something private for your team, open registry vs curated marketplace vs private registry is the shorter read.

Corrections to this page are welcome at hi@skillmd.com. If we have a figure wrong or have described your directory unfairly, we will fix it and say that we did.