apple-voice-memos-mcp

MCP server for Apple Voice Memos (Memo Vocali) on macOS — with speaker diarization (AssemblyAI), whisper.cpp fallback, auto-saved transcripts and a one-command setup.

Quick start

1. Add to your MCP config

{
  "mcpServers": {
    "apple-voice-memos": {
      "command": "npx",
      "args": ["apple-voice-memos-mcp"],
      "env": {
        "ASSEMBLYAI_API_KEY": "your-assemblyai-key"
      }
    }
  }
}

AssemblyAI is optional but recommended: it provides speaker diarization. Without a key, transcription falls back to whisper.cpp (local, no diarization). You can also put the key in ~/Library/apple-voice-memos-mcp/config.json instead of env: {"assemblyai": {"apiKey": "..."}}

2. First run — auto setup

npx apple-voice-memos-mcp --setup

The setup:

1. Installs the bundled app — VoiceMemosMCP.app ships inside the npm package (pre-built, pre-signed, self-contained: no compilation, no Homebrew node dependency). It is extracted to ~/Library/apple-voice-memos-mcp/.
2. Registers the LaunchAgent daemon — a small TCP daemon (127.0.0.1:8917) that spawns the MCP server per-connection inside the app, so it inherits the Full Disk Access.
3. Opens the Full Disk Access panel automatically and waits for you — poll every 5s and confirms the moment you grant it.

The only manual step macOS requires: drag VoiceMemosMCP.app into Full Disk Access (~10 seconds, once). This is a hard macOS privacy requirement for Voice Memos data — no MCP can bypass it.

3. Use it

Restart your MCP client (or /reload). Tools:

| Tool | Description | |---|---| | list_voice_memos | list memos (limit/offset/search) | | get_voice_memo | full metadata of one memo | | get_audio | audio file (path or base64) | | get_transcript | Apple's embedded transcript, if present | | transcribe_memo | full transcription with diarization, auto-saved |

Transcription pipeline

transcribe_memo uses the first available provider:

1. AssemblyAI (if ASSEMBLYAI_API_KEY set) → speaker-labeled transcript (utterances with speaker, text, timestamps)
2. whisper.cpp (fallback, local) → plain transcript with timestamps (needs brew install whisper-cpp + model, see below)
3. Apple SFSpeechRecognizer helper (legacy, VOICE_MEMO_TRANSCRIBER_PATH)

Results are auto-saved in the transcripts folder as:

• YYYY-MM-DD - <title> (id-N).md — readable diarized transcript
• YYYY-MM-DD - <title> (id-N).json — structured utterances

Pass "noSave": true to skip saving. Default folder: ~/Documents/VoiceMemosTranscripts (change via VOICE_MEMOS_TRANSCRIPTS_DIR env or config.json → transcripts.dir).

whisper.cpp fallback setup (optional)

brew install whisper-cpp
mkdir -p ~/.cache/whisper.cpp
curl -L -o ~/.cache/whisper.cpp/ggml-small.bin \
  https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.bin

Configuration reference

~/Library/apple-voice-memos-mcp/config.json:

{
  "assemblyai": { "apiKey": "..." },
  "transcripts": { "dir": "~/Documents/VoiceMemosTranscripts" },
  "whisper": { "cliPath": "/opt/homebrew/bin/whisper-cli", "modelPath": "~/.cache/whisper.cpp/ggml-small.bin" }
}

Environment variables (higher precedence): ASSEMBLYAI_API_KEY, VOICE_MEMOS_TRANSCRIPTS_DIR, VOICE_MEMO_TRANSCRIBER_PATH.

Status & troubleshooting

npx apple-voice-memos-mcp --status

• "unable to open database file" → Full Disk Access missing: run npx apple-voice-memos-mcp --setup (it reopens the panel)
• daemon not running → launchctl load ~/Library/LaunchAgents/it.otomatik.voicememosmcp.plist
• after a package update the permission is lost → new app version = new code signature; re-run --setup and re-grant once

Why the daemon + bundled app?

macOS TCC evaluates the responsible process for privacy checks. A plain node child of your MCP client inherits the client's permissions (none for Voice Memos). The LaunchAgent runs the server inside the signed, user-authorized VoiceMemosMCP.app, which makes the permission stick reliably.

Requirements

• macOS 14+ (Voice Memos opened at least once)
• Node.js 18+ (any version — the daemon app ships its own runtime)
• Xcode CLT not required (launcher is prebuilt in the package)

Credits

Server core forked from jwulff/apple-voice-memo-mcp, with:

• Node 26 compatibility fix (better-sqlite3 13, prebuilds)
• TCC/LaunchAgent architecture with bundled pre-signed app
• AssemblyAI diarization + whisper.cpp fallback + transcript auto-save

License

MIT