Developer API

Use Cosmic 1 through the API.

Ultrac speaks the OpenAI chat-completions format from this site's own /v1. Create a key, set the base URL, and use the cosmic-1 model ID. It is the same API the Ultrac agent signs in to, so one key covers your terminal and your code.

Create a key Request access

Quickstart

Send a chat completion request.

The request needs this site's base URL, your key as a bearer token, and model: cosmic-1. Any OpenAI SDK works unchanged.

Base URL

https://ultrac.io/v1

Auth

Authorization: Bearer ultrac_sk_...

Model

cosmic-1

curl https://ultrac.io/v1/chat/completions \
  -H "Authorization: Bearer $ULTRAC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "cosmic-1",
    "messages": [{"role": "user", "content": "Analyze this incident."}]
  }'

Keys are created and revoked on the API keys page once your account is approved. A key is shown once, stored only as a hash, and can be turned off at any time.

Streaming and tool calling work the way they do at OpenAI: send stream: true for server-sent chunks, and tools with tool_choice to have the model call your functions and return tool_calls. The request reference has a worked example.

Work that takes minutes does not fit in one request. Queue it as a task instead: POST /v1/runs answers before the model starts, a worker does the work, and you follow the event log over server-sent events - dropping the connection and rejoining from the last event you saw costs nothing. Long-running tasks has the endpoints and a copyable loop.

Limits

What each plan allows.

Stated per account, for the plan confirmed on it; until one is, Free applies. Every key also has its own per-minute bucket at the same figure, so a 429 names the key that is hot and ten keys cannot multiply the plan. Nothing here is billed.

Plan	API requests / minute	API requests / day	Dashboard / minute
Free	10	200	20
Pro	60	2,000	40
Team(at 3 seats; the daily figure grows per seat)	300	9,000	60
Ultra	1,200	40,000	120

A request over the limit answers 429 with a retry-after header. Full reference, including the request fields and error codes, is in the docs.