Airlock extends LiteLLM's OpenAI-compatible API. The running proxy is the authoritative source for the endpoint schema because its configured models and enabled features determine the live surface.
With a proxy running, open:
/airlock/docsfor Airlock's conceptual API notes and the interactive schema./openapi.jsonfor the OpenAPI document consumed by client tooling.
For a remote proxy, replace the relative URL with that proxy's base URL, for
example https://airlock.example.com/airlock/docs.
Airlock forwards OpenAI-compatible POST /v1/embeddings requests through the
normal authenticated proxy path for explicitly configured embedding aliases.
In the shipped benchmark configuration, use text-embedding-3-small or
openai/text-embedding-3-small. Chat-only aliases, smart, fuzzy aliases, and
unconfigured models receive an OpenAI-shaped 400 with
error.code=model_endpoint_not_supported; Airlock does not route or fall back
an embedding request. For text-embedding-3-small, dimensions must be an
integer from 1 through 1536 and encoding_format must be float or base64.
Completion-only controls are rejected before dispatch with
error.code=invalid_embedding_option.
The live schema includes Airlock additions to the standard endpoints, including client attribution, supported request metadata, 429 responses, and response headers. Use these static pages for the behavior behind those fields:
The OpenAPI document is authoritative for request and response shape. The versioned configuration template is authoritative for aliases and enabled features. This reference and the guide pages explain behavior, operational constraints, and examples that do not fit cleanly in a schema.