Skip to main content
A concurrency limit caps how many of something are in progress at once; a rate limit caps how many start per unit of time. Every limit is counted in Redis, so it holds across all your servers — adding a server does not multiply it. All are off unless you set them, and anything over a limit is refused as rate_limited.

​
Set a project’s limits

An admin key, and only for its own project. The PUT replaces every limit: a field left out, or null, goes back to the platform default (LIMIT_LIVE_CALLS, LIMIT_OUTBOUND_CALLS_PER_SECOND, LIMIT_MESSAGES_PER_MINUTE), which is unlimited unless the deployment set one. The project read (GET /v1/projects/{id}) shows them under limits. A new value applies to the next call; calls already live keep going, and count toward it.

​
Live calls

What happens to a call over the limit depends on who started it:
  • Inbound (Twilio, Telnyx, Asterisk, WhatsApp voice): refused at the carrier before it is answered — busy on Twilio (<Reject reason="busy"/>) and Telnyx (a reject command, cause USER_BUSY), handed back to the dialplan unanswered on Asterisk (its next priority, so the PBX’s own fallback takes it), a rejection on WhatsApp. Answering only to say no would bill minutes, and so would queueing. Your webhook subscribers get call.rejected:
    The call never started, so it has no call record and call_id is null; provider_call_id is the carrier’s id for it.
  • POST /v1/calls/outbound and POST /v1/webrtc/connect: 429 rate_limited, with details: {"limit": "live_calls", "max": 20}. Nothing is dialled and no call record is written, so back off and retry.
A call counts from the moment it is admitted — an outbound call while it rings — until it ends however it ends: a hangup, the agent ending it, an error, a transfer away from the agent. If the server running a call dies, its slot frees itself within two minutes rather than staying taken: a crash costs you capacity briefly, never for good.

​
Outbound calls per second

Carriers cap how fast one number may start calls, and flag a number that dials faster. outbound_calls_per_second holds each of the project’s numbers under its carrier’s cap: a dial past it is 429 rate_limited before anything is dialled, with how long to wait.
The window slides: no more than max dials from one number in any one second, whichever servers placed them. A refused dial does not count against the window, so a retry after retry_after_ms finds room unless another dial took it first.

​
Messages per minute

Every message a text conversation receives costs an LLM turn, so a looping integration or an abusive sender can run up the bill as fast as it can post. messages_per_minute caps the inbound messages in each session — per conversation, so one runaway conversation never holds back the others. A message past it is not stored and never reaches the LLM:
  • Chat API (POST /v1/chat): 429 rate_limited, with details: {"limit": "messages_per_minute", "max": 20, "retry_after_ms": ...}.
  • SMS and WhatsApp: no reply is sent. There is no one to hand an error to, and answering would spend the very thing the limit saves.
If Redis is unreachable, calls and messages are admitted and a warning is logged. A limit is a guardrail; it should not be the thing that takes your phone line down.