Quickstart
From an invite to streamed tokens, without writing any Atlas-specific code.
Before you start
Atlas is invite-only. An Atlas operator creates your organization and sends the invite; accepting it puts you in the console as either an owner or a member. Both roles can create their own API keys. See Organizations and roles.
Accept your invite
The invite email links into the console. Accepting it joins you to the organization that was created for you — there is no public sign-up path, and you cannot create a second organization from inside the product.
Create an API key
In the console, create a key and give it a label describing what it is for. The full key is displayed once, at creation. Atlas stores only an irreversible hash of it — nobody, Atlas operators included, can show it to you again.
atl_live_7k2m9qr4vx8thn3p_wj6b2yd0fs1gzr4mcqvn8ka3hpx91tCopy it into your secret store now. If you lose it, revoke it and create another. See API keys.
Point your client at Atlas
Set the base URL and the key on the OpenAI client you already use.
export ATLAS_API_KEY="atl_live_..."Send a request
Use an Atlas model identifier from the catalog —
atlas-small-1, atlas-mid-1, atlas-large-1, or atlas-code-1.
A non-streaming request
The response is the OpenAI chat.completion object, with one additive field:
{
"id": "chatcmpl-...",
"object": "chat.completion",
"created": 1757376000,
"model": "atlas-mid-1",
"choices": [
{
"index": 0,
"message": { "role": "assistant", "content": "Red, yellow, and blue." },
"finish_reason": "stop"
}
],
"usage": { "prompt_tokens": 21, "completion_tokens": 7, "total_tokens": 28 },
"request_id": "req_01k4v9m2..."
}request_id is additive to the OpenAI shape, so SDKs ignore it. It is also
returned as the x-request-id response header, which is where the OpenAI
clients surface it on exceptions. Quote it when you report a problem — it is
the only handle Atlas has on a specific request, because the content is not
retained.
A streaming request
See Streaming for the frame format and how a failure mid-stream is reported.
Check what it cost
Usage appears in the console, and on the API through a read-scoped key
(atl_read_), which is a separate key from the inference one:
curl "https://api.inference.runatlas.com/v1/usage?group_by=model" \
-H "Authorization: Bearer $ATLAS_READ_KEY"Beta usage is free; the cost figures are what it would cost at the provisional prices. See Usage and cost.
Next
Pick a model
Four rungs — small, mid, large, and a coding model — with their context lengths and capabilities.
Handle the errors
What to retry, what to fix, and why an overload is not a rate limit.
Know the limits
Per-model request and token budgets, and the headers that report them.
Check compatibility
Which request fields Atlas honors, ignores, or refuses.