Provide compute

Browser host

The zero-install way to lend your GPU. Open the host page, pick a model, and your browser starts serving requests over WebGPU.

Requirements

  • A browser with WebGPU — current Chrome, Edge, or other Chromium builds.
  • A real (hardware) GPU adapter. Pied Piper detects and rejects software adapters, since they'd be too slow to serve.
  • Enough VRAM for the model you choose — smaller quantized models need less.

Start hosting

  1. 1
    Open the host page

    Go to /host. Pied Piper probes your GPU and picks the best available adapter. If only a software adapter is found, it tells you why it can't host.

  2. 2
    Choose a model

    Pied Piper suggests a sensible default for your hardware and lets you switch. Loading streams a progress bar as weights download and compile into GPU memory.

  3. 3
    Go online

    Flip the toggle. Your session opens, heartbeats, and starts accepting matched jobs. The page shows live stats — jobs served, tokens, and earnings.

  4. 4
    List it (optional)

    Add a title, description, and price to publish to the marketplace so borrowers can find and target you.

How it serves

The model runs entirely in your tab via @mlc-ai/web-llm. When dispatch matches you a request, it arrives over Supabase Realtime, runs locally, and streams tokens straight back. The dispatcher can lazily swap you to a different model it needs, within what your hardware supports.

The tab has to stay open
Browser hosting only runs while the tab is alive — close it and your session drains and ends. To host in the background without a visible tab, use the Chrome extension. To serve big models, use a native node.

Stopping cleanly

Toggling off puts your session into draining — it finishes in-flight jobs, stops accepting new ones, then ends. That way no borrower is cut off mid-generation.