Guide
Agent Skills marketplaces and directories compared
For agent Skills you can trust, start at the source: the official anthropics/skills repo for Anthropic-authored Skills, and the publishing company’s own repo for anything vendor-specific — Stripe, Resend, Supabase, Semgrep, Playwright and many others now ship first-party Skills. Use install-count leaderboards to see what is popular, raw GitHub indexes when you need exhaustive coverage, and a curated directory when you want a ranked answer with the reasoning attached. The distinction that matters is not size; it is whether anyone checked.
Updated August 19, 2026 · first published August 19, 2026
At a glance
Side by side
| Source type | Strength | Weakness | Use it when |
|---|---|---|---|
| Official Anthropic repo | Highest provenance; reference implementations | Only covers what Anthropic ships | You want the canonical Skill for a common job |
| Vendor repos | Written by the API owner; updated with the API | You must know the vendor publishes one | Integrating against a specific product |
| Install-count leaderboards | Real adoption signal; clean install UX | Popularity is one dimension and lags quality | Sanity-checking what people actually run |
| Raw GitHub indexes | Exhaustive; finds the long tail | No curation, sparse metadata, low signal-to-noise | Nothing curated exists for your niche |
| Curated GitHub lists | Human-picked; trusted maintainers | No metadata, no ranking, no per-agent testing | Browsing a domain to see what exists |
| Agent-native registries | First-class install in that agent; publishing is built in | Scoped to one agent’s ecosystem | You run that agent and want its native flow |
| Curated directories | Ranked answers with stated rubric and verification | Narrower catalogue by design | You want a decision, not a list |
Catalogue size is the wrong metric
Several indexes now claim hundreds of thousands to millions of Skills, produced by harvesting every SKILL.md on GitHub. That number counts files, not usable Skills. The bulk of it is internal repo tooling — contributor workflows, release scripts, PR-review helpers that maintainers wrote for their own projects and never intended anyone else to install.
Our own verification work found this repeatedly: searching a well-known security vendor’s GitHub organisation for SKILL.md returns dozens of hits, all of them internal engineering skills, and zero product-facing Skills. An index that counts those as inventory is not wrong about the file existing; it is wrong about what the file is for.
The useful questions are narrower. Does this Skill exist at a real, fetchable path? Who published it? What is its licence? Was the install command tested, and when? Does it work on the agent you actually use? None of those are answered by a bigger number.
Where we fit, stated plainly
This site is a curated directory, which means it is deliberately smaller than the indexes and deliberately more opinionated than the lists. Every entry carries a composite score computed from a published rubric, a provenance tier, a licence, per-agent verified compatibility, and a date we last checked the install actually worked.
The honest cost of that approach: if a Skill exists for your exact niche and nobody has vetted it, a raw index will find it and we will not. The honest benefit: nothing here is listed because a file with the right name showed up in a crawl. We also remove entries — fifteen catalogued Skills have been de-listed after their repos or install commands failed verification, with permanent redirects left behind rather than dead URLs.
A practical routine
Combining sources beats loyalty to one. The sequence that works:
- 1. Check the vendor first
- If you are integrating with a product, look for that company’s own Skill before anything else. Vendor Skills track API changes; third-party advice does not.
- 2. Check the official repo
- For generic jobs — documents, review, design, debugging — an Anthropic-authored Skill is usually the reference answer.
- 3. Use a curated directory to choose
- When several candidates exist, a ranked comparison with stated criteria beats reading six READMEs.
- 4. Fall back to indexes for the long tail
- Niche and new Skills show up in raw indexes first. Read the SKILL.md before installing — see the audit guide below.
- 5. Check your agent’s own registry
- Some agents now ship one. OpenClaw has ClawHub (openclaw skills install <package>, or npm i -g clawhub for the standalone CLI), and it reads SKILL.md natively with the same two required frontmatter fields. An agent-native registry gives you the smoothest install for that agent, at the cost of being scoped to it.
- 6. Re-check before you depend on it
- A Skill is a live dependency. Verify the repo is still maintained and the licence has not changed before it enters a production workflow.
FAQ
Common questions
Which agent Skills marketplace should I use?
More than one, for different jobs. The official Anthropic repo and vendor repos are the highest-trust sources; install-count leaderboards show adoption; raw indexes give exhaustive coverage of the long tail; curated directories give a ranked answer with the reasoning shown. No single source is best at all four.
Are Skills from big indexes safe to install?
Treat them as unreviewed until you review them. A Skill is instructions your agent will follow and may include executable scripts, so read the SKILL.md, check the licence, and check what it wants access to. That is true of curated sources too — curation is a filter, not a guarantee.
How is a curated directory ranked?
On this site, by a published composite score with nine weighted inputs — install count, provenance, stars, recency, verified cross-agent compatibility, documentation depth, install ergonomics, licence, and verification freshness. The rubric and weights are on the methodology page so you can disagree with them specifically.
What about ClawHub and OpenClaw?
ClawHub is OpenClaw’s public registry for skills and plugins, installable either through the agent (openclaw skills install <package>) or as a standalone CLI (npm i -g clawhub). OpenClaw reads SKILL.md natively and requires the same two frontmatter fields, so the format is portable. What is not established is per-skill behaviour: OpenClaw’s docs do not claim Claude Code skills work unchanged, and they explicitly note that Codex CLI’s skills directory is not an OpenClaw skill root. We have not run a compatibility pass, so we do not yet list OpenClaw on individual skill pages.
Do I need a marketplace at all?
No — a Skill is just a folder, so you can clone it straight from a repo, and writing your own is often the right answer for repo-specific procedures. Directories exist to help you choose between the ones that already exist.
Related
Keep reading
- How to audit a Skill before installing it →
The ten-minute read that catches most of what goes wrong.
- Our scoring methodology →
The nine weighted inputs behind every score on this site.
- Best agent Skills by use case →
The top pick for each developer job, with the reasoning.