Limits

Every number on this page is enforced in code. Where a limit has both a default and a ceiling, both are given: the default is what you get if you say nothing, the ceiling is the highest value the platform will accept.

​

Agents

The 256 KB plan limit only bites when you send a whole definition inline — an agent_team or an inline agent on a call request. Referencing an agent_id sends nothing. Builtin web-search tools carry their own option caps: xAI allows 5 domains and 20 X handles; OpenAI allows 100 allowed domains. See reasoning and builtin tools.

​

Tools

​

Inside the code sandbox

A code operation runs on an isolated service, not in the call worker.

​

Agent tasks

See task errors and limits.

​

Knowledge bases

​

Calls

These are platform limits, not settings. None of them is configurable per workspace, per agent or per call. The one ring window you can set is a transfer operation’s ringing_timeout, 5–120 s.

​

Batches

​

Images

See vision and images.

​

Text conversations

​

Observability

Call and conversation lists are not limited to 7 days; that ceiling is on the aggregated charts.

​

Webhooks

Because there is no retry, your endpoint has to be up when the event fires or the event is gone. See webhooks.

​

Secrets and provider keys

​

Retention

​

Other names and identifiers

​

Pagination

Every list endpoint takes limit and offset and returns { items, has_more, limit, offset }. There is no single platform-wide page size — the ceiling is per endpoint: offset has no ceiling. There is no cursor and no total count — has_more is computed by fetching one extra row. See pagination.

​

Two limits that do not exist

Both of these come up in evaluation, so here they are stated plainly rather than left for you to discover.

​

There is no general API rate limit

Nothing throttles inbound requests to the API. There is no per-token, per-workspace or per-IP quota, no 429 from the platform, and no Retry-After header to read. A 429 you see came from one of your providers being relayed, not from us. That is the current state, not a promise about the future. Build a client that behaves reasonably rather than one that relies on being stopped.

​

There are no concurrency, spend or duration caps you can configure

• No concurrency cap. Nothing limits how many calls a workspace runs at once. The only concurrency control in the product is max_concurrency on a call batch or an email batch (1–10), which governs that one batch and nothing else. Two batches at 10 dial 20 at a time.
• No spend cap. There is no budget, no maximum monthly charge and no hard stop. The low_balance_threshold on your credit balance is a warning at $2 and blocks nothing — the only thing that stops new calls is the balance actually reaching zero.
• No configurable call-duration cap. The 3-hour maximum above is platform-wide and cannot be lowered for a workspace, an agent or a call. If you need a call to end sooner, do it in the agent: an end_call operation the model can reach, or a prompt that closes the conversation.