API reference

Set your client's base URL to https://api.arscelare.ai/v1.

Authentication

Send your API key in either form:

Authorization: Bearer <key>
# or
Authorization: Api-Key <key>

A missing, unknown, or revoked key returns 401 authentication_error.

Endpoints

  • GET /v1/models. This public endpoint returns an OpenAI-compatible model list containing ars-1.
  • POST /v1/chat/completions. This endpoint requires an active API key and sufficient balance.

Browser clients can call these two endpoints directly if the app allows connections to Arscelare. Cross-origin requests support the Authorization and Content-Type headers. Use an API key for chat completions and omit cookies. Only share a key with clients you trust.

Chat completion fields

The endpoint supports messages, temperature, max_tokens / max_completion_tokens, stop, stream, top_p, top_k, min_p, presence_penalty, frequency_penalty, repeat_penalty, reasoning_effort, seed, stream_options, and response_format (text, JSON object, or JSON schema). Tool calls, functions, and n values other than 1 are not supported.

The two output-limit fields are aliases. Values must be positive integers no greater than 8,192, and they must match if you send both. If you omit them, Arscelare reserves the full output allowance. Messages are never truncated. The fully rendered prompt plus the output allowance must fit within the 65,536-token context window. Invalid requests return 400 before generation or billing.

Reasoning defaults to medium. Supported effort values are none, low, medium, and xhigh. Omitted sampler fields use the model defaults; values you send take precedence. Effort values guide reasoning depth rather than set fixed token budgets. See Model for both sampler profiles.

Prompt formatting is managed by Arscelare. Client-supplied template arguments and roleplay format controls are not forwarded.

Errors

{
  "error": {
    "message": "Human-readable description",
    "type": "<error_type>"
  }
}
  • 400 invalid_request_error
  • 401 authentication_error
  • 402 insufficient_quota
  • 403 when the upstream blocks a request
  • 429 rate_limit_error
  • 502 server_error

Usage object

{
  "prompt_tokens": 0,
  "completion_tokens": 0,
  "total_tokens": 0,
  "prompt_tokens_details": { "cached_tokens": 0 }
}

cached_tokens reports the prompt prefix reused by the model server and billed at the cached-input rate. completion_tokens includes reasoning tokens. A separate reasoning-token count may be absent. When available, reasoning is returned as reasoning_content on the assistant message or streaming delta.