meet-live-assist-server

meet-live-assist-server

Local bridge server for the Meet Live Assist browser extension. The extension captures a Google Meet or Zoom call; this server is what it talks to, and what your own assistant reads.

npx meet-live-assist-server

It prints an auth token on first run. Paste that into the extension's Options once and you are done.

What it does

• Transcript sink. Every caption line is appended to <data-dir>/<session>.txt, the complete record.

• Wake channel. A second file, <session>.wake, gets only the batches worth waking an assistant for (decisions, blockers, your name, real questions). Tailing that instead of the raw transcript is what keeps a 40-minute call from costing hundreds of assistant turns.

• Advice, board, chat. The assistant posts colour-coded advice, decisions and action items; the panel polls them back.

• Snapshots of a shared screen, for visual context.

• Local speech-to-text through whisper.cpp, fully offline, for calls with no captions.

• Text-to-speech into the call (macOS only, see below).

• An MCP front door. The same package ships meet-live-assist-mcp, a stdio MCP adapter over this API, so an assistant reaches the call through tools instead of shelling out to curl:

  claude mcp add meet-live-assist -- npx -y -p meet-live-assist-server meet-live-assist-mcp
  

  It asks this server where its data dir is, so it needs no configuration. The tool that matters is poll: one call for the transcript batch worth a turn, the panel's state and any pending results, with the read position held here, per assistant.

• A restart is survivable. Live state is snapshotted to <data-dir>/.state/ and reloaded on boot, so bouncing the server mid-call costs a couple of seconds rather than the meeting. Consent (letting the assistant drive a tab, post to the meeting chat) and pause/stop deliberately do not come back: they are answers about one call, and meeting codes repeat.

Nothing leaves your machine

It binds 127.0.0.1 only, has no accounts and no telemetry, and writes to files you own. Every route except /health and /auth-check requires the X-MLA-Token header, so a website you happen to be visiting cannot reach it. Transcripts, snapshots and the stored live state are purged after RETENTION_DAYS (14 by default); the state directory and its files are owner-only.

If your assistant uses a cloud model, whatever you route to it is subject to that provider's terms. That part is your configuration, not this server.

Requirements

Node 20+ is the only hard requirement. The audio features need two more binaries:

	install	without it
ffmpeg	brew install ffmpeg / apt install ffmpeg	no local STT, no audio routing
whisper-cli + a ggml model in ~/.local/share/whisper/	brew install whisper-cpp	no local STT (captions still work)

Text-to-speech is macOS-only. It uses say and afplay. Everywhere else the server runs fine and advice appears in the panel as text; only the spoken output is missing.

Config

All optional, all env vars.

var	default	what it does
PORT	8848	the extension has host permission for 127.0.0.1:8848; changing it needs a matching extension setting
TRANSCRIPTS_DIR	~/meet-live-assist/transcripts	data dir (transcripts, snapshots, .mla-token)
RETENTION_DAYS	14	purge transcripts and snapshots older than this (0 keeps them forever)
WAKE_BASE_MS / WAKE_MAX_MS	10000 / 90000	wake-gate window: starts here, doubles on an empty batch up to the max
WAKE_FORCE_MS	180000	flush whatever is buffered after this long, gate or no gate
WAKE_MIN_GAP_MS	8000	floor between two wakes
WHISPER_MODEL	~/.local/share/whisper/ggml-base.bin	which ggml model to transcribe with
WHISPER_PROMPT	a generic technical word list	bias whisper toward your own product names and ticket prefixes
FFMPEG / WHISPER_CLI / WHISPER_SERVER	resolved from Homebrew, /usr/local, /usr/bin, then PATH	override a binary path
TTS_VOICE	Zosia	macOS voice for spoken advice

Health

curl -s http://127.0.0.1:8848/health
# {"ok":true,"dir":"...","tools":{"ffmpeg":true,"whisper":true,"whisperModel":"ggml-base.bin","blackhole":false}}

tools tells you which optional features are live, so a missing ffmpeg shows up here instead of as a mystery later.