commonroom-recorder#

Record a commonroom call from the command line, for transcribing the meeting afterwards.

The recorder joins a room as an ordinary, visible participant (default name "Recorder") that stays muted the whole time. It receives every other participant's audio and writes one WAV file per participant — so a transcript with speaker attribution needs no diarization — plus the room chat and a machine-readable event log. By default it also posts a one-line chat notice to each participant so everyone knows the meeting is being recorded.

Usage#

No install needed — run it straight from this repo's GitHub Pages tarball (nothing is published to npm):

npx https://concept-collection.github.io/commonroom-recorder/commonroom-recorder.tgz record <room> [options]

(npx github:concept-collection/commonroom-recorder works too, but builds from source on first run.) Note that npx caches the download per URL — after a new release, use the versioned tarball listed at the landing page or rm -rf ~/.npm/_npx. Or from a clone:

npm install
npm run build
node dist/cli.js record <room> [options]

Options:

--name <name>      Display name in the room (default: Recorder)
--out <dir>        Output directory (default: ./recordings/<room>-<timestamp>)
--duration <sec>   Stop automatically after this many seconds
--notice <text>    Chat line sent to each participant on connect
                   (default: "🔴 This meeting is being recorded.")
--no-notice        Don't send any recording notice
--transcribe       Transcribe on the fly (needs faster-whisper): the
                   transcript grows in <out>/transcript.md during the meeting
--model <m>        Whisper model for --transcribe (default: small)
--language <xx>    Force a language (default: auto-detect)

Stop with Ctrl-C. Requires Node >= 22 (built-in WebSocket). The WebRTC stack is @roamhq/wrtc, which ships prebuilt binaries for Linux and macOS.

Transcribing#

With record --transcribe, transcription happens live during the meeting: a single faster-whisper model stays loaded, audio is transcribed in chunks cut at natural pauses, and <out>/transcript.md is continuously rewritten — open it (or watch cat) while the meeting runs, and it is finalized the moment you stop. Text typically appears within ~8 seconds of a pause (someone speaking non-stop can lag up to ~30 s until the force-cut). If the transcriber ever fails, the recording is unaffected.

Alternatively (or to redo a recording with a bigger model), transcribe afterwards:

npx https://concept-collection.github.io/commonroom-recorder/commonroom-recorder.tgz transcribe <recording-dir>

Both produce transcript.md — a merged, speaker-attributed transcript of the meeting with the chat and join/left events interleaved on one timeline:

**[10:00:02] Alice:** So the agenda today...

> [10:00:15] 💬 **Alice:** here's the doc link

*[10:00:19] — Bob joined*

**[10:00:20] Bob:** Sorry I'm late...

(plus transcript.json with the same items, structured). Speech recognition runs locally through whichever engine is found (--engine to force one): faster-whisper (pip install faster-whisper), whisper.cpp's whisper-cli (pass the ggml model file via --model), or the openai-whisper CLI. --model defaults to small; --language forces a language instead of auto-detecting. Raw per-file ASR output is cached in <dir>/asr/, so re-running is instant — use --force to re-transcribe.

Output#

<out>/
  audio/<name>-<peer8>-segN.wav   one file per participant per connection
                                  (48 kHz mono s16 PCM, typically)
  chat.txt                        human-readable chat + join/left log
  events.jsonl                    every event with ISO timestamps: join, left,
                                  chat, mute/unmute, segment start/end
  manifest.json                   session summary: room, participants,
                                  segments with start/end times and durations
  inbox/                          drop a file here to send its content to the
                                  room as a chat message (see below)
  AGENT.md                        instructions for a monitoring AI agent

Everything is written incrementally (tail -f chat.txt works live; the manifest is rewritten at every segment boundary and every 30 s), so a crash loses at most about a second of audio. A file only starts when a participant's first real audio arrives — someone who never unmutes produces no file. If a participant disconnects and returns, they get a new numbered segment; the manifest's per-segment start times let a transcript interleave speakers on one timeline. During a segment, silence is padded by wall clock, so a sample's position in the file always tracks elapsed time.

The transcribe subcommand consumes exactly these files; because each WAV is silence-padded to track wall-clock time, an ASR timestamp within a segment plus the segment's startedAt is already the meeting timeline.

How it works#

The p2p layer is commonroom's, ported to Node: the same nostr presence/signaling topics (knowing the room name IS the key), the same schnorr-signed events (with a fresh ephemeral keypair per run), the same deterministic-initiator WebRTC mesh and control data channel (hello, mute notices, chat, bye). To the browsers in the room the recorder is indistinguishable from a participant whose mic and camera are muted — it counts toward the room cap of 8 and appears in the participant list.

Two deliberate deviations from the browser client:

• Receive-only media. The video m-line is negotiated sendonly from the recorder's side (a placeholder track that never produces a frame), so no video is ever sent to the recorder — with up to 7 participants that saves several Mbit/s and all the decode CPU. Audio is symmetric (a silent placeholder goes out, like any muted mic).
• Files instead of tiles. Each remote audio track feeds an RTCAudioSink whose PCM goes straight to an incrementally-written WAV.

Letting an AI agent join the conversation#

The recording directory doubles as an interface for an independent agent (or anything else) that monitors the meeting and occasionally says something:

• Follow the meeting by re-reading transcript.md (with --transcribe), chat.txt, or events.jsonl — they all grow live.
• Interject by writing a file into inbox/: the file's whole content is sent to the room as one chat message, appearing under the recorder's display name, and the file is deleted once sent. Write atomically (create as *.tmp or a dotfile, then rename); messages are capped at 2000 characters.
• AGENT.md, written into every recording directory, contains ready-made instructions for the agent — point it there. The default guidance: only interject when it clearly helps, keep it to a sentence or two, and any task-specific instructions it was given take precedence.

Testing#

npm run test:loopback runs an end-to-end test with no browser: it starts the recorder and a synthetic participant that "speaks" a 440 Hz sine and sends a chat message, then verifies the WAV really contains the tone and the chat made it to disk. It uses the real public nostr relays, so it needs network access. For a real-world test, run the recorder and join the same room at https://concept-collection.github.io/commonroom/ from a browser.