Yap

Yap MCP: private, on-device speech to text for AI assistants

Give Claude, Cursor, VS Code, Windsurf or any MCP client the ability to transcribe audio files on your own computer, and to search what you dictated with the Yap app. No upload, no API key, no account.

Yap MCP runs the same speech engine as the Yap dictation app: NVIDIA Parakeet TDT v3 through sherpa-onnx, on your CPU. Audio never leaves your machine; the only network request is a one-time, checksum-verified download of the speech model (about 670 MB) from Hugging Face.

Tools

Ask your assistant things like "transcribe ~/Downloads/interview.m4a with timestamps", "summarize the voice memo on my desktop", or "what did I dictate about the budget this week?".

Install

Requires Node.js 20 or newer (22.13+ for search_dictations). macOS (Apple Silicon and Intel), Windows x64 and Linux x64/arm64.

Claude Desktop, one click: download yap.mcpb and open it (macOS and Windows). Or add to claude_desktop_config.json:

{
  "mcpServers": {
    "yap": { "command": "npx", "args": ["-y", "github:founden-bot/yap-mcp"] }
  }
}

Claude Code:

claude mcp add yap -- npx -y github:founden-bot/yap-mcp

Claude Code plugin: /plugin marketplace add founden-bot/yap-mcp, then /plugin install yap@yap. The plugin runs the server from its own folder (node ${CLAUDE_PLUGIN_ROOT}/src/index.js) with the dependencies pinned in package-lock.json, and adds the transcribe-recordings skill.

Gemini CLI: gemini extensions install https://github.com/founden-bot/yap-mcp.

Cursor, Windsurf, VS Code, Cline, Zed and other clients: add a stdio server with command npx and arguments -y github:founden-bot/yap-mcp.

The first transcription downloads the model once (a few minutes on a typical connection) and answers with the progress while it does; after that, everything runs offline. Transcription runs about ten times faster than real time on a modern laptop, so an hour of audio takes about six minutes.

In the terminal

The same engine as a command: npx -y -p github:founden-bot/yap-mcp yap-transcribe memo.m4a prints the transcript (add --timestamps for start times, several files at once are fine).

Languages

English, German, French, Spanish, Italian, Portuguese, Dutch, Polish, Russian, Ukrainian, Swedish, Danish, Finnish, Czech, Slovak, Hungarian, Romanian, Bulgarian, Greek, Croatian, Slovenian, Latvian, Lithuanian, Estonian and Maltese.

Privacy

• Audio is read from the path you give, decoded in memory and discarded. Nothing is uploaded, logged or stored.
• search_dictations opens the Yap history database read-only and returns only the dictations that match; that text then goes to the assistant you are using, like anything else you show it.
• Settings: YAP_MCP_CACHE moves the model cache (default: ~/Library/Caches/yap-mcp on macOS, %LOCALAPPDATA%\yap-mcp on Windows, ~/.cache/yap-mcp on Linux). On Windows, a model already downloaded by the Yap app is reused.

Dictate instead of type

Yap MCP handles recordings. To speak into any app as you work (your editor, Slack, email, the Claude and ChatGPT apps), use Yap: hold a key, talk, and the text appears where your cursor is, transcribed on your computer.

License

MIT for this code. The speech model is NVIDIA Parakeet TDT 0.6B v3 (CC-BY-4.0), ONNX conversion by the sherpa-onnx project (Apache-2.0).

Made by Yap, O-mega Enterprise Inc. Questions: hello@yapapp.cc