AI voice agents inside web and mobile apps, over a WebSocket

How it is built

1. 01
   POST /v1/agents

   Create the agent: prompt, opening line, knowledge base and a voice engine. It is the same agent a phone call would reach.

2. 02
   POST /v1/realtime/sessions

   From your server, with your API key: name the agent, optionally the allowed browser origins and your own metadata, and get back a client_secret that works once, for one minute, for that agent.

3. 03
   wss://api.phone.wixzel.com/v1/realtime

   Your client opens the socket with the secret and streams audio. In a browser, RealtimeSession from wixzel-phone/realtime handles the microphone, playback and barge-in; in Flutter, RealtimeConnection from the wixzel_phone package speaks the protocol and your app supplies the microphone and speaker.

4. 04
   GET /v1/calls/{id}

   Every session is a call record with channel web: transcript, duration, cost and your metadata, and the same callCompleted webhook when it ends.

The engine to use

What it costs

Prepaid: add credit from $5 and the sessions spend it. Estimate your own mix.

Frequently asked questions

Can I put a Wixzel Phone voice agent in my website or mobile app?

Yes. Your server calls POST /v1/realtime/sessions with your API key and receives a client_secret; your page or app opens wss://api.phone.wixzel.com/v1/realtime with it and streams audio both ways. The agent, engine, prompt and knowledge base are the same ones a phone call would use. No SIP trunk and no phone number are involved.

Does my browser or app need my API key?

No, and it must never have it. The client only ever holds a client_secret minted by your server, which works once, for one minute, for one agent. For browsers, pass allowed_origins when you mint it and the socket is refused from any other site, so a secret copied out of your page is useless elsewhere.

Which SDKs support the realtime API?

The TypeScript SDK (wixzel-phone on npm) ships a browser client, RealtimeSession, that asks for the microphone, encodes and plays audio and handles the user interrupting the agent. The Dart SDK (wixzel_phone on pub.dev) ships RealtimeConnection for Flutter and Dart, which speaks the protocol while your app supplies the microphone and speaker. Anything else uses the protocol directly: JSON frames over a WebSocket, with audio as base64 G.711 µ-law at 8 kHz.

What audio quality do web and app sessions have?

Phone quality. Audio is G.711 µ-law at 8 kHz in both directions, the same as a phone line, because the sessions run through the same media path as calls. That is fine for speech and a real limitation next to platforms whose browser audio is wideband; it is stated here rather than discovered later.

How is a web or app session billed?

Exactly like a phone call on the same engine: the same per-minute price, from the same prepaid balance, per second, itemised in GET /v1/usage/events. There is no telephony cost because there is no phone line. A session counts toward the account's concurrent-call limit, five by default. Minting a session costs nothing, so a user who opens your page and denies the microphone costs you nothing.

How long can a session last?

Ten minutes by default. Set max_duration_seconds when you mint the session, anywhere from 10 seconds to an hour, and the server ends the session at that point. The agent can also end it, and so can your client.

Does it work in Flutter?

Yes. The Dart SDK's RealtimeConnection handles the socket and the protocol. Your app records 16-bit PCM at 8 kHz with any recording plugin, converts it to µ-law with the helper the SDK provides, and plays the agent's audio back with any PCM player. Turn on echo cancellation, or the agent hears itself through the speaker.

Other uses