Whisper

You run a lightweight MCP server that leverages Whisper to transcribe audio locally on your machine. It integrates with clients to provide fast, on-device transcription using multiple Whisper models, supporting various audio formats and optional timestamps.

How to use

Set up and run the Whisper MCP Server, then connect your MCP client to start transcribing audio locally.

How to install

Prerequisites you need before installation.

• macOS (Apple Silicon tested)
• Node.js 18+
• whisper-cpp via Homebrew: brew install whisper-cpp
• ffmpeg via Homebrew: brew install ffmpeg

Install the Whisper MCP Server globally or run it directly using your preferred method.

npm install -g whisper-mcp

Or run directly without a global install.

npx whisper-mcp

Configuration

Configure the MCP client to connect to Whisper MCP Server using an MCP JSON file. You can run the server through npx or locally via Node when developing from source.

{
  "mcpServers": {
    "whisper-mcp": {
      "command": "npx",
      "args": ["-y", "whisper-mcp"]
    }
  }
}

If you are developing from source locally, you can run the server with Node from the built distribution path.

{
  "mcpServers": {
    "whisper-mcp": {
      "command": "node",
      "args": ["/path/to/whisper-mcp/dist/index.js"]
    }
  }
}

Apple Voice Memos workflow (optional)

For a complete voice memo workflow, you can run Whisper MCP alongside Apple Voice Memos MCP. This pairing gives you a seamless path from memo capture to transcription.

{
  "mcpServers": {
    "apple-voice-memo-mcp": {
      "command": "npx",
      "args": ["-y", "apple-voice-memo-mcp"]
    },
    "whisper-mcp": {
      "command": "npx",
      "args": ["-y", "whisper-mcp"]
    }
  }
}

Tools and usage patterns

This server exposes a compact set of tools to transcribe audio and manage Whisper models. You use these tools through your MCP client to perform on-device transcription and model management.

Notes and tips

If you are using a local development setup, you can point the client to your local built server path and keep iterating without publishing the package.

You can choose from multiple Whisper models (tiny, base, small, medium, large) to balance speed and accuracy. Models are stored locally (default path shown in documentation). If you need to switch models, you can configure the tool call to specify the desired model.

Available tools

transcribe_audio

Transcribe an audio file using Whisper. Parameters include file_path for the absolute path to the audio file, model to use (tiny.en, base.en, small.en, medium.en, large), language code, and output_format (text, timestamps, or json).

list_whisper_models

List available Whisper models and their download status, including model name, size, download status, and local path.

download_whisper_model

Download a Whisper model for local use. Specify the model to download (tiny.en, base.en, small.en, medium.en, large).