AI Voice Agent
Hosted voice agent, AI compute included.
STT + LLM + TTS + telephony. One bill. Sub-second time-to-first-audio.
/ Programmable voice CPaaS
Per-second billing. Prepaid wallet. Hard $50/day fraud cap on day one. Card-on-file = $5 credit, never a charge.
Four products. One platform.
Hosted voice agent, AI compute included.
STT + LLM + TTS + telephony. One bill. Sub-second time-to-first-audio.
BYO carrier voice, in and out.
Plug your own Asterisk / FreeSWITCH / LiveKit / Pipecat into our SIP edge.
Buy and route US DIDs in seconds.
Search by area code, attach to a trunk or AI agent, port-out anytime.
Programmable A2P messaging.
US 10DLC and Toll-Free, pass-through carrier fees published transparently.
REST + Webhooks
POST a call. Pass an agent_id and we run STT, model, and voice — billed per second. Pass a trunk_id and bridge to your own SIP endpoint instead.
// Node.js (server-side)
import Pstn from 'pstn-node';
const pstn = new Pstn(process.env.PSTN_API_KEY);
const call = await pstn.calls.create({
to: '+15558675309',
from: '+14155552671',
agent: 'agent_01HX4P5Q9KZ2',
metadata: { lead_id: 'lead_2912' },
});
console.log(call.id, call.status);
// → call_01HX4PKQX9V2T6HMEWFYZGTX21 queuedLatency budget
Every voice agent is a stack of vendors with their own clocks. Here's ours, broken out so you can audit it.
Carrier ingress → STT first partial
Deepgram Nova-3 streaming, Hillsboro edge.
~180ms p50
STT final → LLM first token
Groq Llama 3.3 70B. No caching tricks.
~140ms p50
LLM first token → TTS first audio out
Cartesia Sonic Turbo, streaming chunks.
~210ms p50
Time-to-first-audio target: ≤ 640ms p50 · ≤ 900ms p95 on a real PSTN handset. Measured weekly. Status published at status.pstn.dev.
$5 on signup
Card-on-file is how we keep burner-email farms out — not how we sneak a first charge in. Auto-recharge is off by default. Hard $0 cutoff when the wallet hits zero.
REST, webhooks, HMAC-signed CDRs. No carrier negotiation. No 60-day net-terms. Just an API key.