Anyone in the room with a Deepgram API key can turn on a live transcript from the new panel in the control bar. A mesh call already delivers every participant's audio to every other one, so a single browser can transcribe the whole room: it opens a separate streaming connection per speaker, which makes attribution structural rather than inferred, with no diarization to misread. Deepgram's pre-recorded REST endpoint is cheaper per minute but sends no CORS headers, so a browser cannot call it without a proxy; the streaming WebSocket connects directly, which is what keeps the page serverless. Browsers cannot set an Authorization header on a WebSocket either, so the key rides the subprotocol as ['token', key], Deepgram's documented client-side scheme. Note that a failed WebSocket handshake hides its HTTP status, so a rejected key and an unreachable network are indistinguishable from the client; the stream gives up after two never-opened sockets rather than reconnecting into a wall on every utterance. Billing follows the audio sent, so only speech is sent. A voice-activity gate on each participant's audio streams while someone is talking and holds the socket open through the silence with unbilled KeepAlive messages, which is what keeps an hour-long call from being billed as eight hours of room tone. A 320 ms pre-roll buffer means opening the gate does not clip the first word, and the panel reports the seconds actually sent so the cost is visible while it accrues. The gate's noise floor adapts only while it is shut, is seeded from the first frame and held shut for a second at the start, and rises by at most a factor of four per step; without the first a long utterance would raise the bar against itself, without the second a participant in a noisy room would open the gate on their first frame and never close it, and without the third a slammed door would deafen it. The transcript is deliberately not sent to the other participants. Text attributed to someone but relayed by someone else is text they cannot vouch for, which is the same objection that keeps chat history from being replayed. What is shared is the fact that it is running, as a badge on the transcriber's tile and a line in the chat, and it is not a room setting: nobody else can turn it on or off. The key stays in the browser it was typed into and has nothing shareable derived from it. Unlike the chat, the transcript persists: it is stored under the room it belongs to and restored on entry, with buttons to copy it, save it as text, or clear it, the last of which confirms first. This does leave meeting transcripts in local storage, which the README says plainly. The gate and the store are the two pieces where a quiet mistake is expensive, so both have node checks; see the Testing section of CLAUDE.md for the bundle-then-run recipe.