Build an embedded voice agent

Build the integration in two places: a Pipecat client in your app and a brain WebSocket in your backend. Voqalize connects them for each call and runs the voice path between them.

The three pieces

• A brain. Your code. Subclass Brain, implement on_user_message.
• An agent. A record on our side: a name and your brain’s URL.
• A client. Stock pipecat on your page. You write no transport code.

Recommended path

1. How a session works — follow one call from your app to the brain and back.
2. Quickstart — run the smallest complete web example.
3. Connect your app — choose browser or backend session creation and connect a Pipecat client.
4. Build the brain — add speech, screen actions, tools, context and conversation history.
5. Use another agent framework — connect an existing framework through the text-and-actions wire.
6. Deploy the brain — use an inbound WebSocket in production or the outbound relay when the environment cannot accept ingress.
7. Test the brain — run protocol scenarios without a microphone or live model.

Use keys and authentication when choosing where a session may be created. Add the avatar after the call path works.

What Voqalize runs

Voqalize runs WebRTC, recognition, synthesis, endpointing, turn-taking, interruption and optional recording. Your app uses Pipecat’s client transport; your brain receives finalized text and returns speech and actions.

Once the complete call works, continue with improving the agent and operating calls.