API reference - Abstract

Item	Value
Base URL	`https://api.abstract.inc`
Authentication	`Authorization: Bearer` with your API key for inference; WorkOS sign-in for CLI credentials
Request and answer bodies	JSON, server-sent events when `stream` is `true`
Version	`Abstract-Version: 2026-10-27`, optional; see Versions and request ids
Request id	`Request-Id` on every answer

These pages do not send requests. Run requests from your own machine, with the key in your program’s environment:

curl -s https://api.abstract.inc/v1/models -H "Authorization: Bearer $(abstract token)"

Every error has the same body. See Errors and limits.

Versions and request ids

Send Abstract-Version to pin the API version your client was written for. 2026-10-27 is the only version. A request without the header is answered under the newest version. A version the API does not answer is refused with 400 unsupported_api_version before anything runs, so nothing is billed. Every answer names the version it was made under in its own Abstract-Version header.

curl -s https://api.abstract.inc/v1/models \
  -H "Authorization: Bearer $(abstract token)" \
  -H "Abstract-Version: 2026-10-27" -D - -o /dev/null

Every answer from a /v1/ route carries a Request-Id header: req_ and 32 hex characters, new for each HTTP attempt, a retry included. An error repeats it in its body as error.request_id, and a robot route’s error also as status.request_id. Quote it when you report a problem. A robot answer also carries X-Trace-Id, which names the reply rather than the attempt: a replayed answer keeps its trace. See Robot action API.

Chat completions

The LLM models.

Robot actions

End-effector and native joint action models.

Catalog

The LLM model ids.

Chat completions

Endpoint	Serves
`POST /v1/chat/completions`	An LLM model: the OpenAI Chat Completions API.
`GET /v1/chat/completions/{completion_id}`	An LLM completion your account created
`GET /v1/models`	The LLM model ids

The model in the body names an LLM model: family-version, with an optional variant (glm-5.3, glm-5.3-flash). A robot model id (origin/model) is refused with use_get_action.

Robot actions

Endpoint	Serves
`POST /v1/get_action`	One observation in, one chunk in the selected model’s action space out
`POST /v1/end_session`	Ends an episode before its idle expiry. Not billed.

Use GET /v1/environments/models for the available robot IDs. YAM models and Xiaomi’s native ARX X5 checkpoint return end-effector chunks. FLUX’s DROID checkpoint takes native Franka joint state and returns actions.joint_chunk with action_space: "joint_position". Their camera roles, gripper conventions and cadences differ; follow the selected model’s request contract.

CLI credentials

abstract login signs in through WorkOS and creates a key directly through the API. abstract logout refreshes the saved sign-in and revokes that key.

Endpoint	Authentication	Result
`POST /v1/api-keys`	WorkOS sign-in bearer token	Creates a key; accepts an optional `name` of up to 80 characters
`DELETE /v1/api-keys/{key_id}`	WorkOS sign-in bearer token	Revokes a key belonging to the signed-in account

API keys and browser cookies do not authenticate these credential endpoints. Creation requires a payment method or available account credit and returns api_key, key_id and base_url. The secret is returned only once. Revocation returns revoked: true, or false if the account has no key with that id. All credential responses disable caching. Use the CLI to manage sign-in tokens rather than copying them by hand.