# Limits and backoff.

Wait when told to wait. Ask when a permission or spending decision is needed.

Updated: 2026-09-21.
## Request limits

This environment publishes a key run-request limit of 60 per minute, a public discovery limit of 30 per minute, and up to 20 discovery results per query.

Provider limits and capacity also apply. A `429` response identifies the relevant scope and supplies retry information when available. Respect `Retry-After` and `retry_after_ms`; use bounded exponential backoff with jitter rather than a tight loop. A published limit is a ceiling, not a throughput guarantee.

## Free tools

The backend publishes 50 free runs per day for an unfunded workspace and 1000 for a funded workspace. A zero-price tool still needs an authorised workspace.

## Spending restrictions

Optional key caps and allowed-tool restrictions are controlled by the workspace. New keys do not have optional caps automatically. An exhausted balance or spending cap calls for a human decision, not a new key or retry loop.

## Uncertain execution

If a request might already have reached a provider, retry with the same idempotency key or follow its run ID. A timeout does not establish that no resources were consumed. See [errors](/docs/errors) and [idempotency](/docs/idempotency).

---

[HTML page](/docs/limits) · [Agent index](/llms.txt)