Confirm permission
Server-held identity, spend limits, destination policy, and readiness, without claiming SIP OPTIONS proves PSTN reachability.
FOR HUMANS / YOUR AGENT, YOUR PERMISSION
Imagine asking an assistant to book a restaurant. It should confirm the details, show you exactly what it is authorised to do, make one bounded call, and tell you what actually happened. This is a product design preview, not a live booking system.
ONE TASK, TWO EXPERIENCES
Humans give permission and receive an understandable outcome. Agents get a machine-readable mandate and telecom events. One does not impersonate the other.
“Call Example Restaurant and book a table for two this Saturday, between 7 and 8:30 PM.”
Before calling, the agent should establish which restaurant, the date, the party size and the name to use. If anything material is missing, it asks.
Illustrative permission card only. No active approve button, payment, or call exists on this preview.
THE CONVERSATION
The final product must work whether the other end is a person, IVR, or autonomous AI. Dial-Tone should handle phone signalling; the requesting agent handles conversation and intent.
Server-held identity, spend limits, destination policy, and readiness, without claiming SIP OPTIONS proves PSTN reachability.
Answer detection, interruptions, hold, DTMF, voicemail and call completion are distinct events, not guessed by a model.
Did the network connect? What did the agent hear? Did the restaurant offer a booking reference? Was it independently verified?
A clear outcome, any uncertainty, what the call cost, and whether a fresh instruction is needed.
THE PROMISE, NOT YET THE PRODUCT
A real restaurant booking, a successful telephone call, and a model-generated summary are different facts. The public preview does not place calls or process x402 payments.