API reference

Base URLs

There are two, because inference and account management are separate services.

Inference

https://api.inference.runatlas.com/v1

Chat completions and the model catalog. Authenticated with an inference key (atl_live_).

Usage

https://api.inference.runatlas.com/v1

Usage reads and exports. Authenticated with a read-scoped key (atl_read_) or a console session.

Endpoints

Method	Path	Auth
`POST`	`/v1/chat/completions`	Inference key
`GET`	`/v1/models`	Inference key
`GET`	`/v1/models/{model}`	Inference key
`GET`	`/v1/usage`	Read key or session
`GET`	`/v1/usage/export`	Read key or session

Key creation, revocation, membership, and invitations are performed in the console with a signed-in session. They are not part of the key-authenticated API surface.

Versioning

The /v1 prefix versions the HTTP surface. It has not changed and will not change compatibly-breakingly without a new prefix.

Model identifiers are versioned separately and independently — see Model lifecycle. The JSON Schema subset is versioned separately again, as atlas-json-schema-subset-v1.

Conventions

Content typeapplication/json

Both directions, except a streamed completion (text/event-stream) and a CSV export (text/csv).

TimestampsISO 8601 / Unix

Account and usage surfaces use ISO 8601 strings in UTC. The OpenAI-shaped created field on completions and models is a Unix timestamp in seconds, because that is what the OpenAI shape specifies.

Request idx-request-id

Every response carries one, and error bodies repeat it as request_id. Quote it in any report — content is not retained, so it is the only handle on a specific call.

ErrorsOpenAI shape

{"error": {"message", "type", "param", "code", "request_id"}}. See Error codes.

Compatibility

The inference endpoints are OpenAI-compatible: use the OpenAI Python or Node client with the base URL changed. There is no Atlas SDK. See OpenAI compatibility for the precise field matrix.