Build an embedded voice agent

You write a pipecat client in your page and a brain behind one WebSocket in your backend. An agent record joins them — a name and your brain’s URL — and Voqalize runs what is between: WebRTC, recognition, synthesis, endpointing, turn-taking, interruption and optional recording. Your brain receives finalized text and returns speech and actions. Your page writes no transport code.

Start with how a session works, which follows one call the whole way, from your page to your brain and back. Every page here is a step of that path in detail.

• Quickstart — the smallest complete web example, running.
• Connect your app — the session-creation request, both credential paths, and the one line of pipecat glue around it.
• Build the brain — speech, screen actions, tools, context and conversation history.
• Use another agent framework — when the conversation logic already exists somewhere else.
• Deploy the brain — a route Voqalize dials, or a relay your brain dials out to.
• Test the brain — your brain over the real wire, with no microphone and no model.
• Keys and authentication — which key goes where, and where a session may be created.
• The avatar — a talking head in the page, once the call path works.
• Camera and screen — record the user’s camera and screenshare beside the audio, with pipecat’s own calls.

Once a call works end to end, improving the agent is what changes next, and operating calls is how you read one back.