Whisper Notes MCP Server & CLI
MCP server and command-line access to Whisper Notes — on-device speech-to-text for Mac and iPhone. Claude and any MCP client can search your transcript history, read transcripts (plain, timestamped, or speaker-labeled), and transcribe audio or video files.
100% on-device. The server speaks over a local Unix socket (0600 permissions) — no open ports, no network interface, no cloud. Your audio and transcripts never leave your Mac.
The MCP server is built into the Whisper Notes Mac app — there is nothing to npx or pip install. This repository is its documentation and registry home. Canonical docs live at whispernotes.app/cli-mcp (agent-readable markdown: whispernotes.app/cli-mcp.md).
Set up
- Download Whisper Notes for Mac 1.5.5 or later — Direct Download (DMG) version (the Mac App Store build does not include the Local API). macOS 14+, Apple Silicon.
- Open the app → Settings → Integrations → toggle on Local CLI & MCP.
- Click Install CLI in the same section — this puts
whispernotes (and the wn shorthand, when that name is free) in your terminal.
Connect Claude or any MCP client
Claude Code — one line:
claude mcp add whispernotes -- whispernotes mcp
Claude Desktop — add this to ~/Library/Application Support/Claude/claude_desktop_config.json, then restart Claude Desktop:
{
"mcpServers": {
"whispernotes": {
"command": "/usr/local/bin/whispernotes",
"args": ["mcp"]
}
}
}
Use the full path from which whispernotes — GUI apps don't read your shell's PATH. Any other local MCP client (Cursor, etc.) works the same way: command whispernotes, args ["mcp"]. Transport is stdio; there is no HTTP endpoint.
Then just ask:
- "Read my last five transcripts and tell me what I should follow up on."
- "Find the meeting where we discussed the invoice, and draft a reply email."
- "Transcribe ~/Desktop/podcast.mp3 and pull out every action item."
| Tool | Parameters | Returns |
|---|
list_transcripts | limit (default 20) · source | Recent transcripts with ids, dates, titles |
search_transcripts | query (required) · limit | Full-text matches with snippets |
get_transcript | id (required) · timestamps · speakers · max_chars · offset | One transcript's text |
get_latest_transcript | timestamps · speakers | Your most recent transcript |
transcribe_file | path (required) | Transcribes an audio or video file, saves it to history |
list_models | — | Installed models and which one is active |
Long recordings are paged: when a transcript exceeds max_chars, the response ends with an explicit truncation marker telling the agent which offset to continue from — hour-long meetings never overflow a context window.
CLI
The same binary is a full command-line tool:
$ whispernotes status
WhisperNotes 1.5.5 (340) · SenseVoice Small · ready · 3081 notes
$ whispernotes list --limit 3
9f3c21a8 2026-08-01 09:14 0:42 meeting Weekly sync — pricing review
b2d84e07 2026-07-31 18:03 12:26 recorded Idea dump on the walk home
5a1f9c33 2026-07-30 10:52 47:10 imported interview-with-anna.m4a
$ whispernotes show latest | ollama run llama3 "Summarize the key points:"
$ whispernotes export --format md --audio --output-dir ~/Transcripts
✓ exported 312/312 transcript(s), 312 audio file(s) → /Users/you/Transcripts
| Command | Options | What it does |
|---|
status | — | App version, active model, readiness, note count |
list | --limit N · --source T · --json | Recent transcripts, newest first |
search <query> | --limit N · --json | Full-text search across your history, with snippets |
show [id or latest] | --timestamps · --speakers · --max-chars N · --offset N · --json | Print one transcript (defaults to latest) |
transcribe <file> | --json | Transcribe an audio or video file, save it to history |
export | --format txt or md · --output-dir DIR · --source T · --limit N · --audio · --speakers | Batch-export transcripts, optionally copying the audio alongside |
models | list · select <id> | List or switch models: parakeet, sensevoice, whisper-small, whisper-large-v3-turbo |
language | list · select <code> | List or switch transcription language — choices follow the active model |
mcp | — | Run the MCP server on stdio |
version · help | — | CLI version · usage |
Results go to stdout, progress and errors to stderr — safe to pipe or redirect. --json always emits a single JSON object or array. whispernotes help prints the full usage; a misused command replies with its exact usage line. wn is a convenience alias — scripts and agents should call whispernotes, which always exists.
Privacy & security
- Transport is a Unix domain socket at
~/Library/Application Support/WhisperNotes/api.sock with 0600 permissions — only your macOS user account can connect. No TCP port, no network interface, unreachable from any other machine.
- Off by default: nothing is served until you enable the toggle in Settings.
- MCP clients like Claude ask for your approval before each tool call.
- Transcription runs on-device (Parakeet V3, Whisper, SenseVoice on the Neural Engine / MLX); results are stored only in your local history.
Links
License
The contents of this repository (documentation and metadata) are MIT-licensed. Whisper Notes itself — the application, including the embedded MCP server and CLI — is commercial software; see whispernotes.app for terms.