---
title: Give your team a voice room your mind listens to
date: 2026-11-13
summary: Members talk in a browser tab. The mind writes down each line, makes no model call, and joins in only when someone says its name.
detail: An hour of talk, no model call, until you say "iterate".
---

With this, your mind can hear what your team says out loud, keep it searchable, and answer when someone says its name.

**Why.** We talk through a lot of our work. Until now the mind only heard what someone typed afterwards.

**How it works.** Open the room in a tab and press **Join with microphone**. Each tab connects to every other tab over WebRTC, a small mesh for two to four people; the tabs find each other through short-lived events on `/rooms/main`. Each tab cuts its own microphone into utterances (one ends after 2.5 seconds of silence) and posts each to the mind. The mind transcribes it, drops noise, and keeps the line as said by the member who is signed in, never by what the page claims.

Every line goes to the room's agent as context only. A line that names the mind wakes it:

```js
// any other line: context, no model call
{ llmRequestPolicy: "dont-trigger-request" }
```

Our files:

- `apps/room/index.html` and `apps/room/public/app.js`: the page: the mesh, the microphone, the cutting into utterances.
- `apps/room/app.ts`: `POST /utterance`, the transcription, and the room's events.
- `room.md`: the agent's guide: when to speak, and how briefly.
- `worker.ts`: the mount for members, the case that hands each utterance to `/agents/room/main`, and the one that copies its answers back into the room.

**What it needs.** An OpenAI API key in `/secrets/openai` for `gpt-4o-transcribe`, billed per minute of audio. The room uses public STUN servers and no TURN server, so two people behind strict firewalls cannot connect.

**What surprised us.** The hard part of a listening agent is not listening. It is staying out of the way until it is asked.
