vgpu-shaders

vgpu-shaders — should this page have a shader, and what does that cost?

vgpu is Vercel Labs' WebGPU library "designed for agents": one shader renders in the browser, headless in Node, to a PNG at any resolution, to video, and in CI. WGSL becomes a module graph you import like TypeScript — the loader resolves it, reflects bindings, drops unused declarations and emits compact source at build time.

⛔ First: this skill does not document the vgpu API, on purpose

The API is 258 symbols across 20 packages, and Vercel ships a first-party skill generated from the source docs, stamped with vgpuVersion and a gitSha in its own frontmatter. Restating it here would duplicate an artefact that regenerates itself, and would be wrong within a release.

npx skills add vercel-labs/vgpu                      # the generated API skill
npx -y add-mcp https://vgpu.sh/api/mcp -g            # or the MCP
npx -y vgpu docs find <query> | grep -i <term> | cat <symbol>

⚠️ Their skill is named vgpu; this one is vgpu-shaders precisely so both can be installed.

What follows is the half nobody else writes: whether to reach for it, and what you owe the user once you do.

It is not only graphics — but for a dev-flow project it almost always is

WebGPU is a GPU compute API of which rendering is one use, and vgpu exposes both. Compute, ComputeOptions, StorageBuffer and PingPongStorage are first-class, and initFromDevice(device) adopts a GPUDevice another library already created — the documented case being ONNX Runtime Web's WebGPU execution provider, so a model's output stays on the GPU and a shader consumes it with gpu.device.wrapBuffer(output.gpuBuffer), zero copies. It runs in Node too, with Dawn supplying WebGPU. Ownership is explicit: vgpu never calls destroy() on a device it did not request.

Know that path exists; then note it is almost never ours. In this skill set the model runs server-side, through eve and the AI Gateway — client-side inference is a different architecture with its own bill (model download weight, WebGPU availability, a device the page now owns). It earns consideration only when the data genuinely must not leave the machine and the client is a browser. [VERIFY] the ML matrix against the first-party skill (vgpu docs cat ml) before pricing that; it pins specific webgpu, onnxruntime-web and Dawn versions.