Agents
Give an agent a voice, a brief, and a first line. It listens, thinks, and answers — and it stops the moment someone talks over it.
The reply streams from the language model, and each finished sentence is spoken while the next is still being generated. Waiting for a whole response before speaking puts the entire generation time in front of the first word.
Barge-in cuts the audio locally first, then tells the server — so it goes quiet the instant you speak, not a round trip later. The model stops generating, and the agent remembers what it actually said rather than what it had planned to say.
An agent speaks with a cloned voice profile. Putting one on a phone call requires a profile you have confirmed is your own — recorded, not just ticked.
Phone calls
Outbound calling is written and tested. What does not exist yet is the carrier account, the number registration, and per-customer verification — so it returns a clear error rather than a phone call. We would rather say that here than have you discover it.
Every call opens by saying it is an AI using a synthetic voice. Generated server-side; the agent cannot shorten or skip it.
A call can only use a voice profile the user confirmed is theirs.
Agent audio carries an AudioSeal watermark with no off switch.
A number must be added before it can be called, with a daily limit. There is no bulk-dial endpoint anywhere in the code.
Every attempt is recorded before a carrier is contacted, refusals included. Nothing deletes them.
What the law requires and what is still missing, in public, before launch.
Placing AI voice calls is regulated — the FCC treats a synthetic voice as artificial under the TCPA, Texas requires disclosure inside 30 seconds, and the EU AI Act's open-source exemption does not cover its transparency duties. The safeguards above exist because of that, not in spite of it.