AllocateGateway
One endpoint for every model, metered from the first token.
The Gateway is the front door of your Private Inference Cloud:
https://api.allocate.network/v1, speaking the standard chat completions
wire. Any SDK that accepts a base URL works unchanged.
The hot path
Authentication runs against an edge cache of your key hashes, never a
database. Gateway overhead is held to a 30ms p50 budget, enforced by a
conformance suite that runs against production on every deploy, and
self-reported on every response in the x-allocate-overhead-ms header.
Every response also carries:
| Header | What it tells you |
|---|---|
x-allocate-request-id | The id for this request. Quote it in support conversations. |
x-allocate-overhead-ms | Gateway overhead for this request in milliseconds. |
x-allocate-model | The model that actually served the request. |
No silent fallback
Failures surface as failures: they are metered with finish=error, captured
with the upstream body preserved, and alerted on. When a provider errors, the
error you receive keeps the provider's original response under error.native,
so you debug against what actually happened. The only fallback that ever
fires is the one you declared on a Route.
What rides on it
Every request through the Gateway is metered as it happens and captured into your tenant's experience database for RecursiveDB. One request, three primitives fed.
See the API reference for the full contract.