The skills.sh registry passed one million agent skills and nearly 280 million installs in seven months, according to a report published by Vercel, which runs it. For scale, GitHub needed 27 months to reach a million repositories, Apple's App Store took a little over five years to reach a million apps, and npm took more than nine years to reach a million packages. One caveat belongs up front: these are figures for what developers published and installed on a public registry, not Vercel's own usage or revenue.
Key takeaways
- skills.sh reached one million listings in seven months, against 27 months for GitHub to hit a million repositories and more than nine years for npm.
- Installs are extremely concentrated: 375 skills β 0.04 percent of the registry β account for 62 percent of all installs, the top 1.2 percent account for 94 percent, and close to half of all skills have been installed exactly once.
- Supply skews technical while demand does not; business operations skills draw 74 percent more installs per listing than average, whereas the average software engineering skill draws about 30 percent fewer.
What a skill is, and why the catalogue filled so fast
A skill is reusable instruction for a particular job: the steps to follow, what a good result looks like, how to check it. The model and its tools already supply the underlying capability, so the file only has to carry the judgment specific to the task, and it can be plain language rather than code. Anthropic introduced Agent Skills in October 2025; Vercel opened skills.sh three months later.
The growth rate follows from who can participate. Writing a skill requires understanding a job, not programming a computer, which is a far larger pool of authors than any previous software registry drew on. Once written, a skill installs across agents without being rewritten for each. The report's own framing is that skills pair a much bigger creator base with software's capacity for reuse β and agents themselves now help write the skills, so the catalogue partly builds itself.
The gap between what people publish and what they install
Listings and installs disagree about what agents are for. Technical work β software engineering, agent workflows, data, infrastructure, security β fills more than half the catalogue, a quarter of it software engineering alone, yet demand refuses to concentrate anywhere: software engineering tops the install chart at just 18 percent, ahead of agent workflows at 15 percent and business operations and writing at roughly 11 percent apiece.
Correcting for how many listings each category has turns that gap into a ratio. Per listing, business operations skills are installed 74 percent above the registry average, writing and documents 50 percent above, cloud and infrastructure 42 percent above; software engineering sits about 30 percent below, education and productivity 39 percent below, research 55 percent below. Read plainly, the categories with the fewest authors have the most demand pooling into each file, which is the shape of a market whose supply side is still recruiting from the wrong room. The registry grew out of developer tooling, and the authorship still shows it.
The category worth watching is agent workflows and automation, at 14.8 percent of installs in the classified dataset and the largest category among listings with at least 100,000 installs. These teach an agent to plan, route work, use tools and drive browsers β abilities that apply to whatever job the agent picks up next, unlike a contract-review skill that applies to contracts. People are using agents to improve agents.
A power law, but not a winner
Install activity is concentrated to a degree that would look pathological in a normal marketplace. Nearly half of all skills were installed exactly once. At the other end, 375 listings β four hundredths of one percent of the registry β carry 62 percent of installs, and the top 1.2 percent carry 94 percent. Yet no single skill dominates: even the most-installed one accounts for less than 1 percent of all installs.
That combination is the structurally interesting part. Skills compete inside a job and accumulate across jobs. A team that has chosen an expense-report skill has no reason to add a second one, but every reason to add skills for spreadsheets, research and slides. Each job produces its own winner, and the winners collectively hold the distribution.
The half-installed-once tail is easy to misread as failure. A registry that anyone can publish to in plain language will accumulate skills written for one team, one client or one afternoon, and those were never competing for installs. What the tail measures is how cheap authorship has become, not how much waste the catalogue carries.
Portability decides the rest. Sorted by whether the work is tied to an industry, cross-industry skills hold 66 percent of the classified catalogue and 87.5 percent of installs, at 3.6 times the installs per listing of industry-specific ones. Tidying a spreadsheet or shipping a website is the same job at a bank as at a hospital, and the skills that survive that translation are the ones that get reused.
What changes next
Vercel's argument for the second million is that the first taught agents what everyone already knows, and the value now moves to what only one company knows β when a customer gets a refund, what can ship without another review. The measure should shift with it. Install count is today's proxy for quality, and it is a popularity metric; as models improve, a skill only justifies itself if the agent does the job better with it than without, which points toward tests and benchmarks attached to skills rather than download badges.
The platform layer is converging on the same primitive. Workflow tools are shipping skills as a first-class object alongside tools and memory β n8n's new Agents include reusable, shareable skill files β which is what a registry needs to matter beyond one vendor's AI agent.
FAQ
Do these numbers represent Vercel's own business?
No. They are aggregate counters from the skills.sh registry Vercel operates, describing what third-party authors published and what users installed. They are not Vercel revenue, traffic or product usage figures.
How reliable are the install counts?
Vercel notes that install figures come from aggregate registry counters and do not represent unique people or necessarily independent choices. Catalogue totals count unique listings, and the function and industry breakdowns describe a classified sample rather than the full registry.
What kind of skill gets installed most?
Portable ones. Cross-industry skills account for 87.5 percent of installs and 3.6 times the installs per listing of industry-specific skills, and agent workflow skills β which teach planning, routing and tool use rather than one domain task β are the largest category among listings above 100,000 installs.
