A reproducible snapshot of the public Agent Skills ecosystem on GitHub. Collected 20 October 2026.
Published by groclawa · data collected 2026-10-20
This is a reproducible snapshot of the public Agent Skills ecosystem on GitHub. It exists because the question we get most often isn't "how do I write a skill" — it's "which one should I install?" That's a discovery problem, and discovery problems need numbers.
Cite this page if you use the figures: groclawa, "The State of Agent Skills — 2026 Q3 snapshot", 20 October 2026, groclawa.com/research/state-of-agent-skills.
| Measure | Value |
|---|---|
Repositories tagged agent-skills | 36,028 |
Repositories tagged claude-skills | 12,548 |
Repositories tagged claude-code | 101,543 |
Repositories whose metadata mentions SKILL.md | 4,332 |
Stars across the top 20 agent-skills repos | 1,867,707 |
agent-skills repos created since 2026-07-20 | 22,576 (63%) |
SKILL.md repos created since 2026-07-20 | 1,993 (46%) |
The last two rows are the story. Nearly two-thirds of the tagged ecosystem appeared in a single quarter. Supply is not the constraint anymore; selection is.
/search/repositories), unauthenticated, 20 results per query, sorted by stars.topic:agent-skills, topic:claude-skills, topic:claude-code, "SKILL.md", plus created:>2026-07-20 variants for the growth figures.total_count from the API, which counts repositories matching the query — not individual skills inside them. A monorepo with 200 skills counts once.top-agent-skill-repos.csv in this folder.Limits. Topic tags are self-assigned, so the population is noisy in both directions: repos with skills but no tag are missed, and repos with the tag but no SKILL.md are included. Treat the numbers as an order-of-magnitude signal, not a census.
| Repo | Stars | What it is |
|---|---|---|
morluto/rea | 335,933 | Reverse-engineering agent toolkit |
anthropics/skills | 182,216 | The official Agent Skills repository |
DietrichGebert/ponytail | 168,030 | "Laziest senior dev" reasoning skill |
addyosmani/agent-skills | 108,601 | Production engineering skills |
paperclipai/paperclip | 104,689 | Agent management app |
nexu-io/open-design | 102,510 | Collaborative design agent workspace |
Leonxlnx/taste-skill | 99,326 | Art direction and reference skills |
tt-a1i/archify | 85,828 | Idea/codebase → interactive diagram |
Two patterns worth noting:
1. A catalog is not a moat; curation is. With 36k repos, "we have 400 skills" is not a value proposition. "We tested 400 and kept 60" is.
2. Install counts beat star counts for buyers. Stars measure developer attention. A buyer wants to know whether the thing runs on their agent version. Publish both, and publish the failure cases.
3. Comparison pages will outperform list pages. In a crowded category, "X vs Y" and "best for Z" queries carry the intent. Lists get skimmed; comparisons get read.
4. The long tail is agent-specific. Skills that work in Claude Code may not load in Windsurf, and the folder layout differs. Pages that answer "does this work in my agent" convert far better than generic skill pages.
curl -s "https://api.github.com/search/repositories?q=topic:agent-skills&sort=stars&per_page=20"
curl -s "https://api.github.com/search/repositories?q=topic:agent-skills+created:%3E2026-07-20&per_page=1"
Re-run quarterly. The delta between snapshots is more interesting than any single number.
There are more than 36,000 repositories tagged agent-skills on GitHub, and 63% of them were created after July 2026 (groclawa, State of Agent Skills, Oct 2026).
The official anthropics/skills repository has 182k stars, but community collections outrank it — discovery happens in lists, not documentation (groclawa, Oct 2026).
The 60-row repo table behind this report is available as CSV. Cite the figures with the date and URL above.