Whispernotes

What the whispernotes MCP server does

The whispernotes MCP server gives MCP-compatible agents access to transcript data managed by the Whisper Notes Mac app. It can return recent transcript records, search the full local history, read a selected transcript, retrieve the newest transcript, transcribe an audio or video file, and report installed speech-recognition models.

The server is intended for workflows that combine speech-to-text with another agent task. For example, an agent can locate a meeting by keywords, read its transcript, and prepare follow-up content. New recordings processed with transcribe_file are added to Whisper Notes history, so they can be found by later searches.

How it works

Whisper Notes performs transcription on the Mac using local models, including Parakeet, Whisper, and SenseVoice. The app stores the resulting transcripts in its local history. The MCP process communicates with the app through a Unix domain socket protected with 0600 permissions; it does not expose a TCP port, HTTP endpoint, or network interface.

MCP clients start the installed command with whispernotes mcp using stdio. Transcript reads support optional timestamps and speaker labels. Large results can be paged: when a response reaches its character limit, the server includes a truncation marker and the next offset to request. This allows agents to process long recordings without requiring the entire transcript in one response.

Setup and configuration

Install Whisper Notes for Mac version 1.5.5 or newer from the direct DMG download. The supported environment is macOS 14 or later on Apple Silicon; the Mac App Store build does not include the Local API needed here. In the app, open Settings, choose Integrations, and enable Local CLI & MCP. Then select Install CLI.

For Claude Code, register the server with claude mcp add whispernotes -- whispernotes mcp. Claude Desktop can use a configuration entry with the installed executable as the command and mcp as its argument. Other local stdio clients, including Cursor, use the same command and argument arrangement. Use the full path returned by which whispernotes when a graphical client cannot find the shell-installed executable.

Tools and capabilities

The whispernotes MCP server exposes these tools:

• list_transcripts returns recent entries, with optional source filtering.
• search_transcripts finds full-text matches and provides snippets.
• get_transcript reads one transcript and accepts options for timestamps, speakers, character limits, and offsets.
• get_latest_transcript retrieves the newest transcript, with timestamp and speaker options.
• transcribe_file processes an audio or video path and saves the result to history.
• list_models reports installed models and identifies the active one.

The corresponding CLI also supports status checks, transcript export, model selection, language selection, JSON output, and direct transcript display. Those commands are available from the installed whispernotes binary but are separate from the MCP tool list.

Limitations and notes

The whispernotes MCP server depends on the Whisper Notes Mac application being installed and configured; the repository documents the embedded server rather than providing a separate package to install with npm or pip. It is local to the Mac user account and is off until the integration toggle is enabled.

The application and its embedded CLI and MCP server are commercial software. The repository's documentation and metadata are MIT-licensed, which does not make the application itself MIT-licensed. MCP clients may request approval before tool calls, depending on the client's behavior.