Realtime Agent — register-call token flow + streamed transcripts
Realtime Agent — register-call token flow + streamed transcripts
The Realtime Agent WebSocket reference now documents the recommended token flow for browser and client-side apps: call the new POST /conversation/register-call endpoint server-side to mint a short-lived, single-use access token (30-second TTL), then open the WebSocket with that token so your API key never reaches the browser. Connecting with a raw API key directly is still fully supported for server-side and trusted clients.
The WebSocket now documents the transcript (final per turn) and transcript.delta (streaming, cumulative) events, which stream live from the pipeline — assistant deltas are paced to the audio the user hears. The corrected event set also fixes the transcript role values (user | assistant) and clarifies input_audio_buffer.commit as a reserved no-op.
A new Browser Voice Cookbook walks through connecting a browser to an agent in three steps with the Agent Web SDK using the token flow.