---
title: "Introducing the Voice Room: talk with people and the mind"
summary: "A member-only WebRTC room whose addressed utterances become durable agent requests, with speech and silence kept distinct."
created: "2026-11-15"
updated: "2026-11-18"
revision: "5"
previous: "2c455f896ee921241bea34029ffe0e790d6c4412"
reviewedThrough: "499865e6c07a5cc1c6a715609915ce4b57177046"
sourceDigest: "8bd47d7aae2e88f23f1c0b173a2cf8d52a8fc44677a8dca8a3a6b940581d7e0f"
---

With this, your mind can join a room where members speak to each other, and act when someone clearly addresses it. It is not a public phone agent and it should not act on every sentence it overhears. The room’s guide makes that distinction explicit: ordinary discussion is context, not authorization.

## Audio travels two ways

`apps/room/public/app.js` joins browsers by WebRTC using ephemeral room events for presence and signalling. Each browser also cuts its own microphone into utterances and uploads a bounded WAV to `POST /utterance`. `apps/room/app.ts` transcribes it and appends a durable `room/utterance-transcribed` event to `/rooms/main`, attributed to the platform-authenticated member. A client-provided speaker name is not accepted as identity.

```ts
case "room":
  return this.auth.require(request) || room.fetch(request, this);
// Durable transcript facts differ from ephemeral WebRTC signalling.
const utteranceType = "room/utterance-transcribed";
```

The server bounds upload size and drops common speech-model noise phrases; the browser caps each utterance at thirty seconds. A mention of the mind’s name can wake the room agent; a short follow-on window lets one request span two utterances. A small debounce gathers words before the model starts. `worker.ts` forwards addressed room events to the fixed room agent and mirrors its plain-text answers back as `room/agent-said`, where the page shows them. `room.md` documents the room’s behavior. Transcripts can also enter the member-only [search](/introducing/search) pipeline; they are not blog content.

## The difficult part is permission to act

The microphone belongs to the signed-in member, but a transcription can still be wrong. The agent must require a clear request, tolerate transcription errors, and remain silent during people thinking aloud. Member authentication protects the room; speech instructions do not turn third-party words into trusted orders. Keep presence and signalling ephemeral while making intentional transcript facts durable. Repeated delivery of an utterance or answer must not duplicate the agent request.

## Dependencies and tests

A similar room needs an iterate mind, the existing agent runtime and model funding, browser microphone permission, WebRTC connectivity and an OpenAI API account for `gpt-4o-transcribe`. Keep the transcription key in the vault, restricted to the provider’s API; no key belongs in browser JavaScript. It can incur transcription and agent usage charges. No phone number or new telephony account is needed just to run the browser room. This implementation uses a small peer mesh and public STUN servers, without TURN, so restrictive networks may prevent members from hearing each other. Report that failure rather than claiming a connected call. Wire the host and event handlers together, not only the HTML. Test two members, permission refusal, silent audio, oversized uploads, split utterances, reconnects and repeated events. Keep transcripts and speaker identities private. [Agents](/introducing/agents) explains the durable conversation side; [the blog](/introducing/blog) publishes only the reusable design, never what people actually said.

## Keep similarly named rooms separate

This recipe describes only `apps/room/` on the `room` host: one room, a durable transcript and text answers from its agent. The `meet` and `meet2` hosts currently use a different app, `apps/meet2/`, while `boiler-room` keeps an earlier implementation and its sessions. The `voice-meet` host is another separate WebRTC-and-voice prototype, not this room. Do not copy their screen-sharing, canvas or per-session behavior into this room’s description. The shared `words.ts` and `words-parse.ts` modules supply vocabulary hints and spelling repairs to voice-meet transcription, and vocabulary instructions to the separate voice service. This room’s upload transcriber does not import them: it still uses its own fixed prompt. Likewise, voice-meet conversations use ordinary mind agents with `AGENTS.md` and `apps/voice-meet/agent.md`; that is not a change to this room’s agent or to Canvas2’s dedicated instructions. To add spoken agent replies here, consume the explicit answer event with a separate speech layer; the existing page displays text, not synthesized speech.
