murmur

macOS only. Turns Apple Voice Memos into markdown transcripts. Local only, stdlib only.

Uses Apple's own transcript when it exists and the memo is English; otherwise transcribes with whisper.cpp.

Install

brew install whisper-cpp ffmpeg
uv tool install git+https://github.com/equanimitech/murmur
murmur setup

murmur setup checks for ffmpeg and whisper-cli, then finds a whisper model or downloads ggml-large-v3-turbo.bin (~1.6 GB) into ~/Library/Application Support/murmur/. If a tool is missing it says which, skips the download and exits 1. It is idempotent. Model lookup order: $MURMUR_MODEL (used as-is when set), murmur's own folder, then a copy Vibe already downloaded, if any. sync and transcribe never download. With no model they exit non-zero and point to murmur setup.

Usage

murmur list [--limit N] [--json]           # date, duration, title, id, has Apple transcript
murmur sync [--out DIR] [--since YYYY-MM-DD] [--force]
murmur transcribe <id | filename | path>    # one transcript to stdout
murmur setup                                # check tools, fetch the model if missing

sync writes DIR/YYYY-MM-DD-HHMM-<slug>.md (default ~/Documents/murmur) with frontmatter: id, date, duration, title, source (apple|whisper), language. It skips memos already written (unless --force), skips audio evicted to iCloud, and holds a lock in DIR, so a file watch can fire it repeatedly.

Claude Code skill

/plugin marketplace add equanimitech/marketplace
/plugin install murmur@equanimitech

Then say "sync my voice memos" or "review my memos" (or /murmur). The skill runs murmur sync, lists transcripts with no reviewed: field, newest first, and summarizes each one in its own language: gist, decisions, action items, ideas, people. On your OK it hands action items to a task tool you have connected and adds reviewed: YYYY-MM-DD to the frontmatter. Transcript content never leaves the session without your consent.

No MCP server is included: the skill uses the CLI and the files directly.

Caveats

• Apple transcripts are English-only. Apple forces an English locale, so PT/FR memos get garbage text. murmur trusts Apple's transcript only when whisper detects en.
• Apple writes its transcript hours or a day after recording. Memos under 2 days old with no Apple transcript are left for a later sync.
• Voice Memos data sits in a protected Group Container. The process running murmur needs Full Disk Access: your terminal for manual runs, the host process for unattended ones. For zenborg that's the zenborg-daemon binary inside zenborg.app (System Settings → Privacy & Security → Full Disk Access).

Unattended sync (zenborg)

A jobs.json entry that runs murmur sync five minutes after Voice Memos stops writing:

{
  "voicememos": {
    "id": "voicememos",
    "name": "voicememos",
    "enabled": true,
    "program": "murmur",
    "args": ["sync"],
    "env": { "PATH": "~/.local/bin:/opt/homebrew/bin:/usr/bin:/bin" },
    "trigger": {
      "kind": "watch",
      "paths": [
        "~/Library/Group Containers/group.com.apple.VoiceMemos.shared/Recordings/CloudRecordings.db",
        "~/Library/Group Containers/group.com.apple.VoiceMemos.shared/Recordings/CloudRecordings.db-wal"
      ],
      "debounceSeconds": 300
    }
  }
}

If your zenborg version doesn't expand ~ or look up bare program names on PATH, write the absolute paths instead.