Today's GitHub trending page reads like a single category: Cloudflare's security-audit skill added more than 2,000 stars in a day, Matt Pocock's skill collection about 1,100, Google's and Anthropic's a few hundred each, plus a browser CLI from Vercel Labs that ships its own skill file. Five repositories, one genre. And the genre now has a storefront.
The shop
skills.sh calls itself "The Agent Skills Directory" and is run by Vercel. Installation is one line — `npx skills add owner/repo` — and the site lists the harnesses it plugs into: Claude Code, Cursor, Codex, GitHub Copilot, Windsurf, Gemini, Cline, Zed, OpenCode and roughly a dozen more. There are packs, topics, an official section, a leaderboard ranked by install count and, relevant later, a tab marked "Security audits".
The tone is an app store's. At the top sits `find-skills` from Vercel Labs at about four million installs, followed by several Matt Pocock skills in the millions and `agent-browser` at 1.2 million. Whether those numbers count real installs, CI runs or repetitions, the page does not say. They are presented as demand, and they read as demand.
What you are actually installing
A skill is a folder of instructions — usually a `SKILL.md` — plus optional scripts the agent loads on demand. That is exactly the problem. A skill is prose that changes an agent's behaviour, bundled with code that runs under the agent's permissions. Installing one hands your agent new instructions and, often enough, new execution paths.
That a security-audit skill is topping the charts this week is both reassuring and telling. Reassuring, because more teams want their codebases examined. Telling, because Cloudflare's own skill lists an OS-enforced sandbox as a requirement — without one it deliberately refuses to execute target code and leaves its leads as `needs_validation`. A tool built to check security has to secure its own environment first. That caution is not the ecosystem's default.
The distribution layer arrived before the trust layer
On Thursday a Cloudsmith blog post called agent skills "the newest dependency in your software supply chain". That is accurate. Skills are installed by URL from a repository; there is no signature, no provenance statement, no advisory database, no lockfile pinning a version. What exists is an install counter and a tab of audits — a start, but not the equivalent of what npm spent years retrofitting.
Social distribution is meanwhile faster than any review. Benji Taylor, X's head of design, published a skill called Stageshow on October 31 that teaches coding agents to build realistic product mockups. It spread through screenshots and "comment and I'll send you the command" mechanics — installation by copy-paste from a social post. That is the actual novelty: skills are no longer only shared developer-to-developer, they are consumed as content.
The counter-movement is forming too. China's iFlytek published SkillHub this week, a self-hosted registry so teams can curate and govern skills internally — an admission that an open public directory is not sufficient as a source of enterprise code.
What follows
Three questions decide whether the directory becomes infrastructure or an incident.
First, do installs get provenance? While `npx skills add` drops someone else's folder into an agent's context unchecked, every social-media recommendation is a potential vector — not theoretically, but as concretely as a typosquatted npm package.
Second, who is allowed to install? In most harnesses the agent can run shell commands itself. An agent that can call `npx skills add` can top up its own instructions. A confirmation prompt on skill installation is the obvious, cheapest mitigation, and it is not yet universal.
Third, does "audit" have teeth? A tab of audits is worth exactly what the audits are worth. If they come from the same place that lists the skill, they need at least stated criteria and a history of who found what, when.
The demand is real: millions of installs, five trending repositories in a day, four large vendors publishing official skills at once. The question is no longer whether agents will follow other people's instructions. They already do. It is whether anyone other than the author stands behind them.

