Skip to main content
The dashboard can talk to an agent through your laptop’s microphone and speakers — no phone number, no telephony provider, no public URL. This is the one capability the dashboard has that the REST API does not, and it is the fastest way to hear whether a prompt, a voice, or a language choice actually works.
The dashboard container runs Next.js in development mode with the source bind-mounted. That is right for local work and wrong for anything user-facing — see Production deployment.

​
What a browser test call is

A live, bidirectional audio session between your browser and the runtime, running the same STT → LLM → TTS pipeline a real caller would hit. Your microphone replaces the phone line; the runtime’s telephony frame serializer is replaced by Pipecat’s protobuf serializer. You reach it from two places, both rendering BrowserCallSession → CallStage:
  • AgentTestModal.tsx — the Test action on a websocket agent’s card on /dashboard. It shows the call stage next to a summary of the agent’s language, STT, TTS, voice, LLM and model.
  • ReviewStep.tsx — the final step of the agent creation wizard. It creates the agent first, then connects to the id that comes back.
The stage shows transport status, a call timer, live turn latency (user stopped speaking → bot started speaking), mic mute, a mic device picker, Pipecat VoiceVisualizer bars, and live captions from usePipecatConversation.

​
Requirements

To make a websocket agent in the wizard, leave the delivery dropdown on its default — “WebSocket — browser test”. Selecting a telephony provider makes it a telephony agent instead. Categories are covered in Agents and agent categories.

​
The audio path

The dashboard uses the official Pipecat packages:
  • @pipecat-ai/client-js — PipecatClient
  • @pipecat-ai/websocket-transport — WebSocketTransport + ProtobufFrameSerializer at 16000 Hz (matches WEBSOCKET_SAMPLE_RATE)
  • @pipecat-ai/client-react — PipecatClientProvider, PipecatClientAudio, mic control, media devices, VoiceVisualizer, usePipecatConversation, transport state, RTVI events
Factory and connect helpers live in frontend/src/lib/pipecat/createBrowserClient.ts. Before connect(), the helper calls createWebCall({ agent_id }) (best-effort) so the CallLog exists and the socket can carry ?call_id=. Live turn latency is measured in the browser: timestamp on UserStoppedSpeaking, then Turn Xms when BotStartedSpeaking fires. Post-call pipeline metrics are still persisted by the runtime observers as before. WebSocket transport only exposes a local MediaStreamTrack, so VoiceVisualizer always uses participantType="local" (bot visualizer would draw nothing). Bars restyle when the agent is speaking.

​
How it connects

One WebSocket, to the same route telephony uses. The runtime dispatches on the agent’s agent_category, so the path does not change:
org_id comes from the stored session, agent_id from the agent being tested. The socket itself carries no handshake message and no auth token: the runtime resolves the agent from the path and loads its config and provider credentials from the API itself. On disconnect / unmount the Pipecat client releases the microphone and closes the socket.

​
Limitations

Browser test calls are recorded. The browser registers a call_type: web CallLog with POST /calls/web before connecting, so a transcript and recording are stored under the same MinIO paths as telephony calls. The captions shown during the call, however, live only in React state and are gone when you close the modal — read the stored transcript through GET /api/v1/calls/{call_id}/transcript.
Other constraints worth knowing:
  • Telephony agents cannot take a browser call. The runtime expects a provider start event on that socket. Use POST /calls/outbound and a real phone instead.
  • No authentication on the socket. Anyone who can reach the runtime port and knows an org_id and agent_id pair can open a session. Do not expose port 7860 publicly without a proxy that authenticates. See Security hardening and Public voice URLs.
  • A test call in the wizard creates a real agent. The Test call step calls POST /agents before it can connect. Abandoning the wizard afterwards leaves the agent in your organisation.