AI Gateway API

The gateway serves two API dialects on the same port. The dialect is chosen by the route; everything behind it — alias resolution, budgets, cache, the upstream call, metering, audit — is dialect-agnostic and operates on one canonical representation.

Client dialect is independent of upstream dialect. An Anthropic-SDK caller can be served by a vLLM upstream and an OpenAI-SDK caller by Anthropic.

Required headers

Every request to the data plane carries the platform tenancy headers:

Header	Required	Purpose
`X-Kis-Tenant`	yes	Tenant id — partitions registry, cache, budgets and audit
`X-Kis-Product`	yes	Product id
`X-Kis-Environment`	yes	Environment id
`X-Kis-Block`	yes	Calling block, for cost attribution
`X-Kis-Customer`	no	Customer id
`X-Kis-Priority`	no	`critical`, `normal` (default) or `background`

A request missing a mandatory header is rejected rather than served against a default tenant.

Model endpoints

OpenAI dialect

POST /v1/chat/completions        chat completion, sync or SSE
POST /v1/embeddings              embeddings
GET  /v1/models                  aliases resolvable by this caller
POST /v1/batches                 submit a batch
GET  /v1/batches/{id}            batch status
POST /v1/batches/{id}/cancel     cancel a batch
GET  /v1/batches/{id}/results    batch results
POST /images/generations         image generation
POST /audio/speech               text to speech
POST /audio/transcriptions       speech to text

Anthropic dialect

POST /v1/messages                message completion, sync or SSE
POST /v1/messages/count_tokens   token count without generating

model names a gateway alias, not a provider model — support.classifier.v3, not claude-haiku-4-5. Which upstream that resolves to is configuration.

{
  "model": "support.classifier.v3",
  "messages": [{ "role": "user", "content": "where is my order?" }]
}

Streaming

Both dialects stream over SSE when the request asks for it. Mid-stream events are recorded, so a replayed streamed call reproduces the stream rather than only the final text. Cancelling mid-stream is supported and recorded as such in the audit trail.

Admin surface

A separate port behind mTLS and platform authentication.

POST   /admin/v1/aliases              create an alias
GET    /admin/v1/aliases              list aliases visible to the tenant
GET    /admin/v1/aliases/{name}       one alias, with resolved inheritance
PUT    /admin/v1/aliases/{name}       publish a new version
DELETE /admin/v1/aliases/{name}       retire a tenant alias

GET    /admin/v1/prompts/{name}       fetch a prompt, any version
PUT    /admin/v1/prompts/{name}       publish a new version

GET    /admin/v1/budgets/{tenant}     current consumption against every scope

GET    /admin/v1/audit/{request_id}   the audit record for one request
GET    /admin/v1/stats                gateway statistics
GET    /admin/v1/health               readiness

Alias objects carry name, dialect, cache (with enabled and mode), budget (including cost_ceiling_usd_per_tenant_per_day), fallback_count and ab_routing. Platform-owned aliases are readable but not writable by a tenant.

Errors

The error body is rendered in the caller’s dialect.

OpenAI envelope:

{
  "error": {
    "message": "human-readable, no customer data",
    "type": "invalid_request_error",
    "code": "canonical_error_code"
  }
}

Anthropic envelope:

{
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "message": "human-readable, no customer data"
  }
}

Note the asymmetry: the OpenAI envelope carries the canonical code, the Anthropic one does not. If you are matching on error identity, use the OpenAI dialect or the request id.

Status	Meaning
`400`	Malformed request
`401` / `403`	Authentication or authorisation failure
`404`	Alias not resolvable for this tenant
`429`	Budget or rate limit exceeded — sets `Retry-After`
`5xx`	All upstreams in the fallback chain failed

429 responses set the Retry-After header in seconds. Messages never contain customer data, so they are safe to log and surface.