Inference API reference

This is the compact route map for hosted inference. For request examples, start with the quickstart and modality pages.

Auth modes

Route family	Raw x402	Compose Key	Notes
`/v1/*`	Yes	Yes	Native Compose inference.
`/external/v1/*`	No	Yes	OpenAI-compatible external clients.
`/.well-known/opencode`	No	No	Public config document.

Model routes

Method	Path	Description
`GET`	`/v1/models`	Generated model catalog.
`GET`	`/v1/models/all`	Extended generated catalog when bundled.
`POST`	`/v1/models/search`	Catalog search and filters.
`GET`	`/v1/models/{model}`	One model row.
`GET`	`/v1/models/{model}/params`	Optional parameter metadata for a model.
`GET`	`/external/v1/models`	OpenAI-shaped model list.
`GET`	`/external/v1/models/{model}`	One OpenAI-shaped model row. Slash-containing IDs are supported.
`GET`	`/v1/families`	Family catalog.
`GET`	`/v1/families/{family}`	One family row.
`GET`	`/v1/modalities`	Modality catalog.
`GET`	`/v1/modalities/{modality}`	One modality row.
`GET`	`/v1/modalities/{modality}/operations`	Operations for a modality.
`GET`	`/v1/modalities/{modality}/operations/{operation}/models`	Models for an operation.

Generation routes

Method	Path	Streaming	Notes
`POST`	`/v1/chat/completions`	Yes	OpenAI Chat Completions shape.
`POST`	`/v1/responses`	Yes	Responses-style input and output.
`GET`	`/v1/responses/{id}`	No	Retrieves a stored response record.
`GET`	`/v1/responses/{id}/input_items`	No	Returns stored input items.
`POST`	`/v1/responses/{id}/cancel`	No	Cancels a response or async job.
`POST`	`/v1/responses/{id}/input_items`	No	Appends input to a live realtime response.
`POST`	`/v1/embeddings`	No	OpenAI embeddings shape.
`POST`	`/v1/images/generations`	Provider-dependent	Image generation.
`POST`	`/v1/images/edits`	No	Image editing.
`POST`	`/v1/audio/speech`	No	Returns audio bytes.
`POST`	`/v1/audio/transcriptions`	No	Returns text or JSON.
`POST`	`/v1/videos/generations`	No	Submits video generation.
`GET`	`/v1/videos/{id}`	No	Polls video job status.
`GET`	`/v1/videos/{id}/stream`	Yes	Streams video job status.

External generation routes mirror the supported OpenAI-compatible subset under /external/v1. The API also registers POST /api/inference as a legacy alias for chat completions.

Request fields

The core normalizer understands standard OpenAI fields and keeps provider-specific pass-through fields in customParams when possible. Common fields:

Field	Routes	Notes
`model`	All generation routes	Public catalog ID.
`messages`	Chat completions	OpenAI chat message array.
`input`	Responses and embeddings	Responses-style content or embedding input.
`tools`	Text routes	Chat uses nested function objects; Responses uses flat function objects.
`tool_choice`	Text routes	`none`, `auto`, `required`, or the protocol’s named function choice shape.
`response_format`	Text routes	`text`, `json_object`, or `json_schema`.
`stream`	Text and streamable media routes	Must be boolean.
`stream_options`	Streaming routes	Preserved; usage inclusion is forced where metering requires it.
`parallel_tool_calls`	Text routes	Catalog-validated. `true` returns `400 unsupported_parallel_tool_calls` unless the compiled model profile declares support.
`reasoning_effort`	Text routes	Passed through when compatible with the selected provider/model.

Response headers

Header	Description
`X-Request-Id`	Request ID minted or accepted by the root middleware.
`X-Receipt`	Native Compose receipt.
`x-routing-primary-model`	Primary model selected by routing policy.
`x-routing-final-model`	Final model after retry policy.
`x-routing-attempts`	Number of provider attempts.
`x-routing-fallback-used`	`false` with the single-target routing policy.
`x-public-model`	External route public model ID.
`x-upstream-provider`	External route provider selected by resolver.
`x-upstream-model`	External route provider wire model.

Error shape

Native and external routes return OpenAI-style errors:

{
  "error": {
    "message": "model is required",
    "type": "invalid_request_error",
    "code": "missing_model"
  }
}

External routes sanitize upstream content-filter messages and redact URLs. Diagnostic provider/model headers remain available when safe. Responses streams use the official semantic lifecycle: response.output_item.added, interleavable response.function_call_arguments.delta, response.function_call_arguments.done, response.output_item.done, and one terminal response.completed, response.incomplete, or response.failed. Output-item id and continuation call_id are distinct and stable.

Overview

External Use

Families

Modalities

Guides

Auth modes

Model routes

Generation routes

Request fields

Response headers

Error shape

​Auth modes

​Model routes

​Generation routes

​Request fields

​Response headers

​Error shape

​Related

Auth modes

Model routes

Generation routes

Request fields

Response headers

Error shape

Related