Meeting Assistant

Native macOS meeting automation for OpenClaw: calendar/window detection → prompt before recording → system audio + microphone recording → local transcription → agent-generated meeting notes.

Architecture

Google Calendar / Window Detector
          ↓
 schedule.json → scheduler_daemon.py
          ↓
 meeting_daemon.py ask_record
          ↓
 notify.py prompt: Start recording?
          ↓
 record_audio.py → AudioDaemon.app (Unix socket)
          ↓
 ScreenCaptureKit system audio + AVFoundation microphone
          ↓
 WAV → transcribe.py → whisper-cli
          ↓
 transcript.txt + summary_request queue
          ↓
 OpenClaw heartbeat agent → final summary.md → user notification

Key Features

• No BlackHole required — captures system audio directly with ScreenCaptureKit.
• Native AudioDaemon — a signed macOS app bundle for stable TCC privacy permissions.
• System audio + microphone — records remote speaker audio and local microphone input.
• Half-duplex mixing — prioritize system audio while others speak; fade to microphone during system silence.
• Meeting detection — detects Zoom, Tencent Meeting, Google Meet, Feishu/Lark, WeChat calls, and browser-based meetings.
• Local transcription — uses whisper-cli from whisper.cpp.
• OpenClaw agent summaries — transcription writes a summary_request into the agent queue; the OpenClaw agent generates the final meeting notes and notifies the user.

Install

Clone the full repository first. setup.sh needs repository files and should not be run via curl | bash.

git clone https://github.com/Nowhitestar/meeting-assistant.git
cd meeting-assistant
bash meeting-assistant/scripts/setup.sh

If you are already inside the skill folder that contains SKILL.md and scripts/, you can run:

bash scripts/setup.sh

The installer will:

1. Check macOS, Homebrew, and Python
2. Install sox, ffmpeg, whisper-cpp, terminal-notifier, gogcli, and cloudflared
3. Create ~/.config/meeting-assistant/config.json
4. Compile the window scanner and guide Accessibility permission setup
5. Build and sign AudioDaemon.app
6. Guide Microphone and Screen & System Audio Recording permissions
7. Optionally configure Google Calendar
8. Install LaunchAgent background services
9. Run basic verification tests

macOS Permissions

If system audio is missing, open:

System Settings → Privacy & Security → Screen & System Audio Recording → enable AudioDaemon.app.

Google Calendar and Privacy

Calendar authorization uses gog; refresh tokens are stored in the system Keychain. The repository and skill package must never contain real client_secret*.json, refresh tokens, API keys, or personal calendar IDs.

Recommended setup:

# 1. In Google Cloud Console, enable Calendar API and create a Desktop OAuth client.
# 2. Download the JSON locally.
# 3. Import credentials and authorize Calendar access.
gog auth credentials ~/Downloads/client_secret_xxx.json
gog auth add your.email@example.com --services calendar
gog auth list
# 4. After auth succeeds, delete the Downloads copy of client_secret_xxx.json.

Then enable Google Calendar in ~/.config/meeting-assistant/config.json:

{
  "calendars": [
    {
      "type": "google",
      "enabled": true,
      "gog_account": "your.email@example.com",
      "watch_calendars": ["primary"]
    }
  ]
}

If credentials were ever posted in chat or committed publicly: delete the OAuth client, revoke the token, create a fresh client, and authorize again.

Configuration

Config path: ~/.config/meeting-assistant/config.json

Key fields:

{
  "audio": {
    "backend": "daemon",
    "output_dir": "/path/to/meeting-assistant/meeting-recordings",
    "silence_threshold": 0.01,
    "silence_duration_sec": 300,
    "half_duplex": true
  },
  "transcription": {
    "whisper_cli": "whisper-cli",
    "local_model_path": "~/Models/whisper/ggml-medium.bin",
    "language": "zh"
  },
  "llm": {
    "enabled": true
  }
}

Notes:

• audio.backend=daemon is the recommended default and does not require BlackHole.
• mic_device / system_audio_device are only for the legacy SoX fallback path.
• If no external LLM command is configured, summaries are delegated to the OpenClaw agent queue.

Scripts

All scripts live under scripts/:

Manual Commands

# AudioDaemon status
echo '{"action":"status"}' | nc -w 2 -U ~/.config/meeting-assistant/audio_daemon.sock

# Manual recording
python3 scripts/record_audio.py start "Test Meeting"
python3 scripts/record_audio.py stop

# Prompt-based recording flow
python3 scripts/meeting_daemon.py ask_record "Test Meeting" "manual-test"

# Transcribe WAV
python3 scripts/transcribe.py /path/to/meeting-recordings/xxx.wav

# Calendar
python3 scripts/check_meetings.py today
python3 scripts/check_meetings.py upcoming
python3 scripts/check_meetings.py week --json

# Window detection
python3 scripts/meeting_detector.py once
python3 scripts/meeting_detector.py daemon

Troubleshooting

AudioDaemon has no system audio

echo '{"action":"status"}' | nc -w 2 -U ~/.config/meeting-assistant/audio_daemon.sock

If sysReady=false:

1. Open System Settings → Privacy & Security → Screen & System Audio Recording
2. Enable AudioDaemon.app
3. Restart AudioDaemon:

pkill -f audio_daemon
open scripts/AudioDaemon.app

WAV exists but is silent

Usually this is a TCC permission issue or the daemon is not ready. Newer versions refuse to start recording when sysReady / micReady is false to avoid fake silent WAVs.

Summary is still a draft

Check the queue:

cat ~/.config/meeting-assistant/agent-queue.json

If there is a summary_request, the OpenClaw heartbeat agent will read the transcript and template, overwrite the final summary, and notify the user.

中文说明

这是一个 macOS 原生会议助手：日历/窗口检测 → 弹窗询问 → 录制系统音频 + 麦克风 → 本地转录 → OpenClaw agent 生成会议纪要。

核心特点：不需要 BlackHole；通过 ScreenCaptureKit 直接捕获系统音频；用 whisper-cli 本地转写；通过 OpenClaw agent queue 生成最终纪要。

快速安装：

git clone https://github.com/Nowhitestar/meeting-assistant.git
cd meeting-assistant
bash meeting-assistant/scripts/setup.sh

Google Calendar 授权使用 gog，refresh token 存在系统 Keychain。不要把 client_secret*.json、refresh token、API key 或个人日历 ID 提交到公开仓库。