~950 ms
One team's dashboards said the agent was fast. Callers heard about 950 ms of delay on every turn.
CallProbe calls your voice agent over the real phone network every few minutes and alerts you when it is slow, silent or broken.
One team's dashboards said the agent was fast. Callers heard about 950 ms of delay on every turn.
A freelancer who fixes production agents keeps finding SIP trunks lost after a restart, and greetings that play before the audio is connected.
Another team's agent took 7 seconds to say hello: 7 seconds of silence at the start of every call.
A real phone call from a Twilio number, on a schedule. It listens, asks one question and hangs up. Vapi, Retell, LiveKit, Pipecat or your own stack: if it has a number, it works.
Greeting latency, response latency, dead air, talk-over, a silent agent, a clipped greeting. Optionally, a transcript checked against the phrases you expect.
Slack when a probe starts failing, when the problem changes and when it recovers. A status page with p50/p95 latency and history.
# synthetic calls with known latencies, then the status page pip install callprobe callprobe demo callprobe analyze callprobe-demo/bookings-latest.wav greeting 2760 ms after answer, lasts 1820 ms response 1 2180 ms after our prompt ended result FAIL - [slow_response] agent answered after 2180 ms (limit 2000 ms) callprobe serve # http://127.0.0.1:8707 # then point it at your real agent (bring your Twilio account) callprobe watch --every 15m
Each probe is a real call: about a minute of phone time plus a minute of your agent's own time.
Each probe uses about 30 seconds of your agent and phone time: one probe every 15 minutes is about 2,880 probes per agent a month. Phone minutes are paid by the caller and are not included. The hosted plans are not live yet: join the waitlist.
We only use your email to contact you about this. Privacy.