Install
npx skills add vercel-labs/vgpuWho it's for
vgpu is for people writing WebGPU in TypeScript: it offers typed shader imports, a small GPU-oriented API, and the same code can run in the browser, headless Node, and tests. The repository also says: an assistant with skill support can install a very thin documentation router that queries the docs bundled with “the vgpu version already installed in the current project.”
A closer fit: you already have, or will have, a TypeScript project. Install the package via Quick Start first, then let the assistant query docs only, or use the CLI for read-only checks. It is not a standalone, version-free API handbook — the skill itself does not include a versioned API reference.
Do not merge the two entries into one. Project dependencies go through pnpm add vgpu. The assistant skill goes through npx skills add vercel-labs/vgpu. In an environment without skill support, do not apply the second command.
First try in 3 minutes
If the current assistant supports skills, the official skill command is npx skills add vercel-labs/vgpu. The repository is https://github.com/vercel-labs/vgpu Follow only that page. The skill installs into the assistant’s skill directory; it does not replace the project dependency.
Prerequisites: the project should already have vgpu installed, because the skill queries the docs bundled with that version. If the package is missing, first run Quick Start in a disposable project: pnpm add vgpu and pnpm add -D @webgpu/types. Do not invent an undocumented installer.
Low-risk first task: do not start by writing a formal scene that submits work to the GPU. After the package is in, run npx vgpu check to validate shaders, or npx vgpu docs to open docs read-only. Success signal: the CLI finishes and returns a check or docs result, not that your business code has been rewritten. npx vgpu examples opens the example gallery; the first time you can look only and not change anything.
If you take the skill path: after install, ask the assistant a read-only question about how this project uses vgpu. Success signal: the answer matches the package you just installed, not a generic handbook unrelated to the project. The docs do not write a slash-command name for the skill, so a natural-language question is enough.
Use cases
Case 1: Install the documentation router for an assistant that supports skills. The command is npx skills add vercel-labs/vgpu. The skill is thin: it has no versioned API reference and instead queries the docs that ship with the vgpu already installed in the project. This fits when you are changing a specific version and worry the handbook will not match the code. If the project has no vgpu yet, install via Quick Start first, then ask. The docs do not write an offline mirror or a fixed docs-site address; vgpu.sh publishes agents.md and llms.txt for models to read, but does not expand publish steps.
Case 2: Develop against the same API across runtimes and validate with the CLI. The public API covers the browser, headless Node (vgpu/node, Dawn backend), and a deterministic mock for tests and CI (vgpu/mock). Day to day you can use npx vgpu docs, npx vgpu examples, and npx vgpu check. This fits when you want to validate shaders before commit, or run tests in CI with the mock. The docs do not write Dawn install details or a CI config template; until you see WebGPU/Dawn ready on this machine, do not assume a browser example will run through in a headless environment. If a command fails, go back to https://github.com/vercel-labs/vgpu and compare.
Limits
Install limit: the project uses pnpm add vgpu and @webgpu/types; the assistant uses npx skills add vercel-labs/vgpu. Without skill support, do not apply the second command. Skill limit: documentation routing only, no versioned API reference; the query target is the version already installed in the project. Runtime limit: browser, vgpu/node, and vgpu/mock share the public API; the docs do not promise identical behavior for every example in all three environments. CLI limit: you can check docs, examples, and check. The only package manager written is pnpm; do not use an undocumented installer.