Five places ship Claude Code skills today: ClaudSkills, Anthropic's official marketplace, the community-run claude-skill-registry, Vercel Labs' agent-skills repo, and the long tail of awesome-style GitHub lists. They overlap on the easy questions (everything is free; everything publishes SKILL.md format) and diverge sharply on the hard ones — how many skills, how curated, who decides what's good. This page lays out the differences and gives a decision tree at the end.
| Catalog | Skill count | Curation | Pricing | Install flow |
|---|---|---|---|---|
| ClaudSkills claudskills.com |
69,000+ | Automated structural admission (score ≥ 50); no editorial gate |
Free; optional Pro $9/mo for QualityScore + 1-click install | Web download of SKILL.md, or 1-click via desktop app (Pro) |
| Anthropic Marketplace github.com/anthropics/skills |
~30 reference skills | Anthropic-authored only; reference + production document skills | Free | /plugin marketplace add anthropics/skills inside Claude Code |
| claude-skill-registry majiayu000/claude-skill-registry |
Several thousand (dedup'd index) | Automated GitHub crawl + community PRs; security scan; no public scoring yet | Free | Go CLI: sk search foo && sk install <path> |
| vercel-labs/agent-skills github.com/vercel-labs/agent-skills |
7 skills | Hand-curated by Vercel maintainers; Vercel-adjacent topics | Free | npx skills add vercel-labs/agent-skills |
| Awesome-* GitHub lists awesome-claude-code, etc. |
Variable; tens to hundreds per list | PR-curated by each list's maintainer; no automated quality gate | Free | Clone or download referenced repos manually |
Two numbers stand out. The catalogs cluster into "hundreds-or-low-thousands hand-curated" (Anthropic, Vercel, awesome lists) and "tens-of-thousands automated" (ClaudSkills, claude-skill-registry). Different bets on the breadth-vs-handpicked axis.
Open registry of SKILL.md files mined twice daily from 26 public sources — GitHub code search, GitHub topics, awesome lists, dev.to, Mastodon, Bluesky, HackerNews, Reddit, Telegram, marketplaces, RSS feeds, and more (full source list at /stats). Every skill that clears a structural-heuristic admission threshold (score ≥ 50) lands in the catalog. Authors never pay anything — no listing fee, no acceleration, no ranking boost. Pro ($9/month, $79/year, or $149 lifetime) adds a multi-signal QualityScore on every skill page in the desktop app plus one-click install. The free tier exposes the full unranked catalog and direct SKILL.md download.
The QualityScore is content-derived only — structural quality of the SKILL.md plus frontmatter depth — with no popularity signals (stars, install counts, social shares). The design rationale is that popularity tracks promotion, not quality, and content-derived scoring keeps the ranking unforgeable by upstream marketing. Full methodology at /about.
Strengths: largest catalog, most mining sources, transparent admission criterion, open dataset access at /data/skills.json. Weaknesses: no human editorial layer between admission and exposure — some admitted skills are mediocre by craft standards even though they pass the structural bar.
The first-party option, at github.com/anthropics/skills. About 30 reference skills organized into Creative & Design, Development & Technical, Enterprise & Communication, and Document Skills. Most are Apache 2.0 open source. The document skills (skills/docx, skills/pdf, skills/pptx, skills/xlsx) are source-available — Anthropic uses them in production but doesn't license them as open source.
Installation is the smoothest in the ecosystem: inside Claude Code, /plugin marketplace add anthropics/skills and Claude handles the rest. On Claude.ai (paid plans) skills can be uploaded via the support flow.
Strengths: officially blessed; quality is uniformly high; install is one command. Weaknesses: small set (Anthropic-authored only); doesn't aggregate community work.
Community-led registry at github.com/majiayu000/claude-skill-registry, fed by automated GitHub crawling plus PR-submitted community entries. Comes with a Go CLI sk (go install github.com/majiayu000/caude-skill-manager@latest) for terminal-native install and a web search UI at majiayu000.github.io/claude-skill-registry-core/. The registry code is MIT; individual skills retain their upstream licenses. A category-quality audit pipeline exists in the repo but quality scores aren't yet exposed in the UI.
Closest direct competitor to ClaudSkills by approach. The biggest practical differences: (1) ClaudSkills mines from 26 sources, this registry from a smaller set focused on GitHub; (2) ClaudSkills publishes a content-derived Pro QualityScore today, claude-skill-registry has scoring planned but not live; (3) ClaudSkills has its own desktop install client, claude-skill-registry uses a Go CLI.
Strengths: developer-friendly CLI, transparent JSON endpoints, MIT pipeline. Weaknesses: smaller catalog, no published quality signal yet, no end-user GUI install.
Vercel's curated set of seven skills at github.com/vercel-labs/agent-skills — vercel-deploy, react-best-practices, web-design-guidelines, react-native-guidelines, react-view-transitions, composition-patterns, vercel-deploy-claimable. Vercel-adjacent topics with a slight web-stack lean. MIT-licensed. Install via npx skills add vercel-labs/agent-skills.
Useful as an example of how a vendor with a specific stack ships skills for their customers. The same shape (vendor-owned, hand-curated, narrowly-scoped) likely emerges from Stripe, Cloudflare, AWS, and other developer platforms as the format matures.
Strengths: officially Vercel-blessed; install is one command; skills are tightly relevant if you're shipping on Vercel. Weaknesses: only seven skills; only valuable for Vercel-focused workflows.
The long tail. Repos like awesome-claude-code, awesome-claude-skills, and similar lists curated by individuals. Each is one developer's PR-curated taste — typically a markdown table of links to other repos containing SKILL.md files. No automated quality signal, no install client, no aggregated browse.
Useful as a starting point for human readers but tough to scale beyond a few hundred entries before the maintainer burns out. Most awesome lists have a six-to-eighteen-month active window before they stall.
The catalogs are complementary, not zero-sum. A typical Claude Code user ends up with skills installed from two or three sources: a few document skills from Anthropic, a stack-specific bundle from their hosting provider, and a long tail of single-skill installs from the open catalog of their choice.
ClaudSkills.com currently indexes 69,000+ SKILL.md files mined from 26 public sources. The next-largest community option is the claude-skill-registry by majiayu000. Anthropic's official marketplace ships about 30 curated reference skills.
No. Every catalog covered here is free to browse and install. ClaudSkills offers an optional Pro tier ($9/month) that surfaces a multi-signal QualityScore and adds one-click install inside the desktop app — but the full unranked catalog and per-skill SKILL.md files remain free for everyone.
On ClaudSkills.com, no — authors never pay anything. Listing fees, acceleration, ranking boost, featured placement, and sponsored Pro Collections are explicitly forbidden by the architecture rules. The same is true of every catalog in this comparison. None of them accept author payments for promotion. If a directory ever offers paid placement, evaluate it against this baseline.
A skill is a single SKILL.md file (markdown plus YAML frontmatter). A plugin is Anthropic's higher-level packaging primitive that can bundle one or more skills along with hooks, settings, slash commands, and MCP server configs. Most skills in every catalog covered here are loose SKILL.md files, not plugins. Anthropic's marketplace also distributes plugin bundles.
ClaudSkills uses an automated structural-heuristic admission threshold — no human curation in the admission path. The claude-skill-registry runs security scanning and taxonomy governance checks but no published quality scoring at the time of writing. Anthropic's marketplace contains Anthropic-authored reference skills only. The vercel-labs/agent-skills repo is hand-curated by Vercel maintainers. Awesome-style GitHub lists are PR-curated by their maintainers.
SKILL.md files, not affiliated with, endorsed by, or sponsored by Anthropic.