The WAVE Dispatch Engine

A prompt is classified in embedding space at the edge and answered with a route, not a completion — which tier should run this work. Milliseconds, and no inference performed to decide.

One routing decision is one metered unit — $0.0001 over x402, or a card subscription. WAVE bills the orchestration layer and never the inference, so local hops we never executed are never token-billed.

Dispatch was the first WAVE service an autonomous agent paid for over HTTP-402. High-frequency, low-stakes decisions are exactly where per-call settlement has to work without friction — so this is where the open path was proven first.

The wall is the product. Keys, prompts, and inference stay on your infrastructure; the edge returns a decision and receives nothing it did not need. BYO-infra and BYO-keys are architectural, not a setting you can forget to switch on.

Run it fully local and it fails closed: if local inference fails, you get an error — not a silent forward of your prompt to a frontier provider. The quiet leak is the failure mode worth engineering against.

Thin clients in five languages, a CLI, an MCP tool surface, and an OpenAI-compatible proxy — point an agent you already run at a new base URL and it routes.

Fast, Expert, Heavy, and Code as explicit local → heavy → frontier chains, so a route resolves to a named policy instead of a single hop. Defined and scaffolded; not yet served by the live decision path.

The proxy logs every request so you can measure your own local-able fraction before enabling offload. Measuring YOUR traffic ships today; a published WAVE benchmark with methodology does not — and we will not quote a savings number until it does.

Voice, captions, transcription, sentiment, and search inherit this plane rather than inventing a parallel one — the same routing policy, the same meter, the same agent path.