Vercel Skills CLI is a command-line installer for Agent Skills: small folders of plain-language instructions that teach a coding AI agent how to perform one specific task. It is published as vercel-labs/skills and run with npx skills, and it is aimed at developers who keep pasting the same review checklist, house style or framework conventions into Claude Code, Cursor or Codex and would rather install those instructions once.
What it does
A skill is a folder with a SKILL.md file inside it: plain-language instructions, plus optional scripts, reference documents and templates. The CLI's job is distribution. You name a skill, it fetches it and writes it where your agent expects to find it.
npx skills add vercel-labs/agent-skillsinstalls a skill from a GitHub repository.- The CLI supports 17 agent platforms, including Claude Code, Cursor, Codex, GitHub Copilot, Gemini CLI, OpenCode and Windsurf, and auto-detects which of them are installed on the machine.
- Other commands cover the rest of a package manager's life cycle:
find,list,use,remove,updateandinit, with a-gflag for global installs. - Vercel also runs skills.sh, a public directory and leaderboard where people publish skills and see install counts.
The scale figures all come from Vercel itself, not from an independent audit: skills.sh lists 1,565,688 published skills, and by Vercel's own eight-week count the most-installed entry is find-skills, with 4.1M installs, whose only job is helping an agent find the next skill to install. The repository itself reports 37.4k stars and 3.2k forks.
How it works
Nothing here is a model or a runtime. A skill is markdown that the agent reads on demand, so the capability you install is a set of written instructions plus whatever scripts ship next to them. That is why the pitch is npm for AI agents: the one-command mental model of installing a package, with a markdown instruction file in the place where code would normally be.
The Agent Skills format is not Vercel's invention. It was created by Anthropic for Claude and then released as a vendor-neutral standard anyone can implement, which is why a single CLI can write skills for agents from several different companies. Vercel announced the combination of CLI and skills.sh directory on January 20, 2026.
Getting started
The repository is github.com/vercel-labs/skills and the code is MIT licensed. The requirements the project names are short: Node.js 22.20 or newer, and a coding agent already installed, because the CLI only writes files into an agent's configuration and does nothing on its own. There is no separate build or release download to hunt for: it runs straight from npm with npx skills. The project does not list operating-system or hardware requirements.
On money: the CLI and the skills in the public directory are free, but the intelligence is not part of the deal. You still pay for whatever agent reads the instructions, through its subscription or an API key, and the materials here do not quote those prices.
When to use it / when not
The main catch is that this is plumbing, not capability. Installing a skill does not make an agent smarter; it moves text onto your disk in the right place, and a directory with more than a million published entries is a count, not a quality filter. The second catch is the Node 22.20 floor, which is newer than the Node many machines carry by default.
Use it instead of pasting the same instructions into Claude when more than one person, project or agent needs the same rules and you want updates you can pull rather than re-copy. Don't use it if you work in a single agent with one or two personal prompts: a file you maintain by hand is simpler than a package manager.
Alternatives
The alternative most readers already use is manual: write the markdown yourself, drop it into each agent's own configuration directory, or just keep pasting instructions into the chat. That works and costs nothing extra. Because the format is open rather than Vercel's property, other vendors can ship their own installers for the same skill folders, so whatever lock-in exists sits with the directory, not with the file format.
Take this repository seriously if you maintain shared conventions for a team that works through coding agents, especially across more than one tool: the format is open, the CLI is a small MIT-licensed utility, and it replaces copying instructions by hand and forgetting to update them with something versioned and installable. If you are a solo user with one agent and a stable prompt, the interesting part is not the CLI but the directory it reads, and you can browse that without installing anything at all.