Rate limits - AtlaSent

Endpoint limits

Endpoint	Limit	Scope
`POST /v1-evaluate` (and `/v1/evaluate`)	100 req/min	Per org
All other endpoints	Plan tier (see table below)	Per org

Plan tiers

Plan	Requests / minute	Burst
Community (free cloud allowance)	100	150
Team	Not yet published — contact sales@atlasent.io	—
Business	1,000	2,000
Enterprise	Negotiated	—

Tier names updated 2026-08-15 to the current commercial model (Community / Team / Enterprise as of 2026-09-27; the free cloud allowance was called Sandbox until 2026-09-26, then Free until 2026-09-27, and its internal plan key is still free). This table previously listed “Pro”; Business is the current purchasable tier that supersedes it (Pro is now grandfathered, non-purchasable) and inherits its published request-rate figures. Team’s own request-rate figures were never published separately in this table and are left unstated here rather than invented — check with sales for the current value.

Response headers

/v1/evaluate responses carry these headers:

Header	Meaning
`X-RateLimit-Limit`	Requests allowed per minute.
`X-RateLimit-Remaining`	Requests left in the current window.
`X-RateLimit-Reset`	Unix seconds when the window resets.

429 response body

A 429 uses the standard error envelope. On /v1/evaluate and /v1-verify-permit it carries retry_after_ms:

{
  "error": "unavailable",
  "message": "Runtime cannot decide right now; retry shortly",
  "status": 429,
  "retry_after_ms": 60000
}

Other endpoints return "error": "rate_limited". When retry_after_ms is present, use it for back-off; otherwise back off exponentially.

Handling 429 responses

Respect the server’s retry hint, then back off exponentially. The SDKs do this for you; if you’re calling the API directly:

import time
from atlasent import AtlaSentClient, RateLimitError

client = AtlaSentClient(...)

def evaluate_with_backoff(req, max_retries=4):
    for attempt in range(max_retries):
        try:
            return client.evaluate(**req)
        except RateLimitError as e:
            if attempt == max_retries - 1:
                raise
            wait_s = e.retry_after if e.retry_after else (2 ** attempt)
            time.sleep(wait_s)

Do not retry 429s indefinitely. For a regulated agent, a sustained 429 should be treated as deny — the agent halts and surfaces a clear error to the user.