vercel-labs/skills

The problem

Coding agents now come in dozens of flavors — Claude Code, Codex, Cursor, OpenCode and many more — and each one has its own place to drop instruction files. If you write a good "how we do code review here" or "how to use this internal library" prompt, distributing it means hand-copying files into each agent's config directory, per project, then repeating that when the instructions change. skills turns that into an install problem instead of a copy problem.

What's different

The usual approach is either one agent's native project rules file (.cursorrules, CLAUDE.md, and friends) or a hand-rolled dotfiles repo with symlinks. Both work for a single agent; neither scales across a team that mixes agents. This CLI adds a discovery layer (skills.sh directory, npx skills find), per-agent installation path detection across a long list of supported agents, and update checking. Notably it also has an experimental path to sync skills shipped inside npm packages, which ties agent instructions to normal dependency versioning — that is the most interesting idea here, and also the least settled one.

Who it's for

Teams that use more than one coding agent, or that maintain internal libraries and want their usage guidance to reach every agent a developer happens to run. It is also reasonable for individuals who want to try community skills without manually editing config. It is not for you if you standardize on a single agent and are happy with its built-in rules file — the extra indirection buys little. Anyone running older Node runtimes should look elsewhere, since it requires Node.js 22.20+.

Cost of adoption

Requires Node.js 22.20 or later; installation is via npx, e.g. npx skills add vercel-labs/agent-skills. The CLI collects anonymous telemetry by default, which must be disabled explicitly with DISABLE_TELEMETRY=1 or DO_NOT_TRACK=1 — expect a review step before rolling it out in regulated environments. Skills are installed into project directories by default, so decide upfront whether you want global installs. The experimental_sync feature is self-described as unstable. MIT licensed, from Vercel Labs, so treat it as a lab product rather than a committed platform.

Verdict

The project is young — created in early 2026 and less than a year old — yet it has already drawn roughly 36k stars and about 3.1k forks, and it showed meaningful monthly momentum on the TypeScript board. That combination says the problem is real and the ecosystem is crowded with early adopters, not that the CLI's command surface has stabilized. The core install/list/update loop is small and low-risk; the npm sync experiment is where behavior may still move. Adopt it if cross-agent skill distribution is an actual pain today, and keep your skill content in ordinary git so you are not locked into the tool's layout.