AIAgentSkills

AI Agent Skills is a directory, not a registry or CDN: it indexes SKILL.md packages published in other people's GitHub repositories (Cloudflare, browser-use, Alibaba, Stripe, individual developers, etc.), reproduces the skill's description and instructions, and links out to the real source for install. It is not affiliated with Anthropic or OpenAI, which it states plainly on its About page. Each skill detail page carries a "Source Code" link to the exact file on GitHub — checked against three: cloudflare/security-audit-skill, browser-use's cloud skill, and affaan-m/everything-claude-code — and all three resolved to live, matching repos.

The site's own About page and /llms.txt list "github.com/bbwdadfg/ai-agent-skills-index" as its "Public skill index." As of 24 September 2026 that repository returns 404 on both the GitHub web UI and the GitHub API, and the owner account bbwdadfg itself now 404s on the GitHub API's /users endpoint too — consistent with the account being deleted or renamed away from that handle, not with the repo being merely renamed. A GitHub-wide search for a repository named ai-agent-skills-index turns up only an unrelated project (tehdarklea/ai-agent-skills-index, which brands itself as powered by a different product, skill.coffee) — not a fork or successor of the missing one. This doesn't affect how the directory itself works: individual skill pages don't depend on that aggregator repo, they link directly to each skill's own upstream source, and those links check out. It does mean one specific citation on the vendor's About/llms.txt pages is currently false and should be corrected or removed by the site.

The two long-form guides — "What Is an AI Agent Skill?" and "How to Install Claude Skills in 2026" — are accurate and specific rather than generic marketing filler: the install guide correctly separates Claude.ai's ZIP-upload flow, Claude Code's filesystem-based personal/project skills, third-party installers (Vercel Labs' npx skills add, GitHub CLI's gh skill install), and Claude Code plugin marketplaces, and explicitly warns that there is no universal "claude install owner-repo" command — a real, useful correction of a mistake other sites repeat. Skill pages reproduce the actual SKILL.md content rather than paraphrasing it, so a reader can judge a skill's real instructions on the page before clicking through to the repo.

The homepage's trust strip advertises "245M Stars" across listed repos as a headline credibility number. But several of the highest-"star" skills on the site come from personal utility repos whose GitHub star counts exceed far more widely used software: affaan-m/ECC shows 266,166 stars and mattpocock/skills shows 268,462 stars on GitHub itself (both verified directly against the GitHub API), more than TensorFlow or Visual Studio Code. That inflation is on GitHub, not manufactured by AI Agent Skills — the site is faithfully displaying real API numbers — but it means the aggregate "245M Stars" figure the site markets as a trust signal is itself built partly on numbers its own FAQ elsewhere admits shouldn't be treated as a quality or security guarantee ("Stars alone are not treated as a security guarantee").

No named operator, company, or founder appears anywhere on the site — About, Privacy Policy, and Terms of Service all route to a single [email protected] address with no legal entity, address, or social profile listed. There's no pricing or advertise page (both 404); per the FAQ, browsing and installing skills is free, and there's no visible route for a maintainer to pay for placement, though "Editor's Choice" and "Trending" sections don't state their selection criteria. The site says skills go through "automated quality and safety checks" before being marked released but does not describe what those checks are; its own advice to installers — review permissions, scripts, and network access before installing anything — is the same advice a reader would need regardless of this site's checks.