Both sites index the same underlying thing: SKILL.md files scattered across GitHub. Anthropic defined the format, and now there are two public directories over it — skills.sh, built by Vercel, and @skills (source), built by us. Same protocol, same raw material, two different answers to the question "how does an agent actually get a skill into context."
The install command vs. the reference
skills.sh's model is npx skills add <owner/repo> — a package-manager metaphor. You pick a skill from its leaderboard (36,110 skills at last count, ranked by install count — vercel-react-best-practices leads at 80.8K), run the command, and it's copied into your project.
@skills's model is a path reference: type @skills:owner/repo/path in your agent's chat and the skill is pulled in for that turn — no copy step, no dependency to keep in sync, no file added to your repo unless you explicitly save it. It works because the format is just a folder an agent can read directly; the install step is optional convenience, not a requirement.
Neither model is strictly better — they're optimizing for different things. An install command gives you a persistent, version-pinned file in your repo, which matters if you want the skill to survive without network access or if your team wants it under code review like any other dependency. A path reference gives you zero footprint and zero staleness — you always get whatever is at that path right now, and nothing lingers in your repo if you only needed it once.
Scale is a real difference, but it's not the whole story
@skills currently indexes 57,720+ skills against skills.sh's 36,110. That's a real gap in raw coverage — but coverage numbers alone don't answer "will I find a good skill for my problem," and neither directory has fully solved discovery at that scale yet (we're mid-fix on our own indexing pipeline right now, more on that below).
skills.sh's real strength isn't the count — it's the leaderboard. Ranking by install count turns "which skill should I use" into a solved problem for the popular cases: if you want React best practices, vercel-react-best-practices at 80.8K installs is an obvious, socially-validated choice. That's a genuinely useful signal we don't have an equivalent for yet.
Where the two philosophies actually diverge
The deeper difference isn't the UI, it's the assumption about what "using a skill" means:
| skills.sh | @skills | |
|---|---|---|
| Primary mechanic | npx skills add — copies the skill into your repo |
@skills:path — references it live, no copy |
| Ranking signal | Install count (leaderboard) | AI Boost rating (does this add capability the model doesn't already have) |
| Footprint | A file lands in your project | None, unless you explicitly save it |
| Backing | Vercel | SylphAI |
| Agent support | Listed per-agent (Claude Code, Cursor, Codex, Copilot, Windsurf, and 13 more) | Any agent that can read a path and has a bash tool |
The AI Boost rating is the one we think matters most for actually deciding whether to use a skill, versus install count which mostly tells you what's popular. A skill can be popular because it's genuinely useful, or because it was first to a common use case — install count doesn't distinguish those. AI Boost is our attempt at asking a narrower question: does this skill add something the model can't already do reliably on its own (a script, a niche API, domain knowledge), or is it mostly restating what a capable model already knows how to do.
What we're doing about the honest gap
We'll say the quiet part out loud: at the time of writing, @skills has close to zero organic search traffic and most of its 57,720+ pages aren't indexed by Google yet. That's a brand-new-domain problem, not a content problem — Search Console shows the sitemap was accepted, the pages just haven't been crawled at that volume yet. skills.sh, with Vercel's existing domain authority and backlink profile, doesn't have that problem.
We're fixing our end of it: deduplicating near-identical entries in the catalogue so we're not spending crawl budget on thin pages, and getting the site linked from places search engines already trust — starting with our own GitHub repo, which already points back here.
Which one should you use
Honestly, try both. If you want a skill copied into your repo with a version you control, npx skills add is the more mature tool for that today. If you want to reference a skill without adding a dependency to your project — for a one-off task, or to try before you commit to installing anything — @skills:owner/repo/path does that with less ceremony.
They're both betting the SKILL.md format wins as the standard. We just disagree about whether "using a skill" should mean installing something or referencing something. Given that both are open — skills.sh's directory and our own repo — you don't have to pick a side to benefit from either.
@skills is open source. Browse the catalogue at atskills.one, or check the reference implementation at github.com/SylphAI-Inc/atskills.