Vercel’s skills.sh registry passed one million published “agent skills” and logged nearly 280 million installs in its first seven months, according to a report the company published on its own blog. A skill is a short, plain-language instruction file that tells an AI agent how a specific person or company wants a job done, the missing context a general-purpose agent does not have on its own.
Anthropic introduced the Agent Skills format in October 2025. Vercel launched skills.sh three months later to host and distribute them, and the new report is Vercel’s own read of its registry’s aggregate usage data, not an independent audit.
Vercel frames the seven-month milestone against three platforms it says took far longer to hit the same round number. GitHub needed 27 months before its repository count crossed one million, by the company’s own account. Apple’s App Store spent more than five years listing its millionth app, and npm took upward of nine years to accumulate a millionth package. Vercel’s explanation is structural: a skill is just a text file describing how a job should be done, so publishing one requires no code, which let the catalog fill faster than platforms built around shipping actual software or apps.
Supply skews toward developer work. Security, infrastructure, data, agent workflows and software engineering together make up more than half of what is listed. Software engineering alone accounts for roughly a quarter of the catalog, per Vercel’s own figures. Demand looks different: software engineering still draws the most installs at 18%, agent workflows follow at 15%, and business operations and writing each pull close to 11%.
Installs per listing point to where demand outstrips supply. Business operations skills pull in 74% more installs than the catalog average, per Vercel’s numbers; a writing or documents skill outperforms by roughly 50%, and a cloud and infrastructure skill by about 42%. Software engineering, education and research run the other way, each landing well under average on a per-listing basis. The pattern suggests developers have already crowded their own shelf, while operations and writing teams remain comparatively underserved.
Install activity is concentrated at the extreme. Nearly half of all one million skills were installed exactly once, while 375 skills, just 0.04% of the registry, account for 62% of all installs, and the top 1.2% account for 94%. Even so, Vercel says no single skill clears 1% of total installs on its own: the market has many small winners rather than one dominant product, because a company that has already picked its expense-report skill has no reason to install a second one but will still add separate skills for spreadsheets or research.
Skills built to work across industries, the kind equally useful at a bank or a restaurant, make up two-thirds of classified listings yet capture 87.5% of installs. Per listing, a cross-industry skill outperforms an industry-specific one by a factor of 3.6, according to Vercel’s breakdown. The company reads this as evidence that portability, whether across jobs or across sectors, is what actually drives adoption.
None of this comes with an independent measure of retention or of whether an installed skill gets used more than once, and Vercel has an obvious interest in the registry it runs looking large and fast-growing. Vercel’s own forecast carries more signal than its adoption numbers. The company expects the standard to shift from popularity to effectiveness: as capable models raise the bar, a skill will have to prove it makes the agent perform the job better, not simply rack up downloads. That points toward skills getting benchmarked the way code libraries already are. Teams standardizing on Agent Skills internally should start logging which installed skills actually change task outcomes now, before that measurement becomes the industry’s baseline instead of a competitive edge.
Reported by Vercel in a blog post on its own site, “State of agent skills.”
