From a sentence to a phone line.
A Callable is a phone line run by an AI agent.
How it works
Say what the line should do.
A plain-language request starts the work.
Answer for my salon, take appointment requests, and take messages after hours
It asks what it can’t infer.
The builder asks one question at a time about what it can’t infer.
- 02Asked back
What language and locale should the agent speak to callers?
It writes exactly two things.
It writes a greeting and a system prompt — the agent’s standing instructions.
First message
Welcome to [Business Name] support. How may I assist you today?
System prompt
- identity_and_scope
- hard_facts
- speech_rules
- unknowns_refusals_escalation
Then it tries to break the draft.
The builder plans caller scenarios: a happy path, a caller it must refuse, and awkward ones in between.
- question-agent-can-answer — PassedPassed
- asks-for-discount — PassedPassed
- vague-about-need — PassedPassed
- changes-subject-mid-answer — PassedPassed
- unrelated-question — PassedPassed
- angry-about-grievance — PassedPassed
- Passed 6/6
Publishing is your act, on the record.
When the transcripts satisfy you, publish a version; the act is recorded.
Draft
your explicit act
Publish
its own dedicated server
Provision
to a number you own
Your number
Your number, answered.
Buy a number in the dashboard, then point it at your Callable.
+1 (___) ___-____
Your Callable
Watch every call it takes.
Edit the draft and publish again; it does not change itself.
| Call type | Transcript | Turn latency | Cost |
|---|---|---|---|
| Pipeline call | Kept | 602–985 ms | at least $0.053/minSoniox + Cartesia + Twilio · price book 2026-08-05 |
| Realtime call · OpenAI | Kept | — | at least $0.013/minTwilio · price book 2026-08-05 |
| Realtime call · Gemini Live | Kept | — | —Not on the price book yet |
| Realtime call · xAI Grok | Kept | — | $0.05/minxAI · price book 2026-08-05 |
Turn latency measured across four real calls over the public phone network, July 2026. Only the pipeline call type is instrumented today.
We measure two things.
Turn latency measured across four real calls over the public phone network, July 2026.
Straight answers about the mechanism.
From the greeting and system prompt you approve before it takes a call. It follows that brief and what the caller has said; it doesn’t go beyond the brief.
No. A Callable is a phone line, not a chat window. It is tested against caller scenarios before anyone can dial it.
The builder plans and grades caller scenarios: a happy path, a caller it must refuse, and awkward cases in between. Tests are text; speech quality is configured, not measured this way.
It says so plainly and takes down what it needs. It does not guess or invent an answer the brief didn’t give it.
On Callable, an AI voice agent is a phone line run by an AI agent. Describe it in one sentence: it asks what it can’t infer, writes the greeting and system prompt, tests caller scenarios, and answers calls only after you publish it.