EarScribe MCP

What EarScribe MCP MCP server does

The EarScribe MCP MCP server makes a small, canonical set of EarScribe information available to MCP-compatible AI clients. Its content covers EarScribe’s local-first transcription workflow, official URLs, support information, and related public guidance. The server is read-only: it does not modify data, upload recordings, or control the browser transcription interface.

EarScribe is the separate browser application that transcribes audio on the user’s device with OpenAI’s Whisper model. It uses WebGPU when available and can fall back to WebAssembly. The application can produce timestamped transcripts and export SRT, VTT, JSON, or plain-text files. Those transcription functions are described by the server’s knowledge surface but are not exposed as MCP operations here.

How it works

The EarScribe MCP MCP server communicates with clients through the MCP stdio transport. Clients start the Node entry point as a local process and receive text or Markdown responses.

Two tools are available:

• get_local_setup returns canonical guidance for using EarScribe’s on-device AI workflow.
• get_official_links returns the canonical EarScribe website, support, and documentation links when available.

The project also documents MCP resources for local setup, FAQs, and links, plus prompts for describing EarScribe and walking through local setup. The listed tools accept no parameters.

Setup and configuration

The project describes two installation paths. Smithery can install it with:

npx -y @smithery/cli install earscribe-mcp --client claude

The client argument can be replaced with cursor, windsurf, or continue. For a source installation, clone the repository, enter its directory, and install dependencies with pnpm install. Configure the client to launch Node with the absolute path to src/index.mjs.

The documented development command is pnpm start, which runs the process over stdio. MCP Inspector can be used with npx @modelcontextprotocol/inspector node src/index.mjs.

Tools and capabilities

The main capability is structured access to information that an AI client can use when answering questions about EarScribe. get_local_setup is appropriate for explaining the browser-based, on-device workflow. get_official_links is appropriate when a user needs the official website, About page, support email, or other canonical URLs.

EarScribe itself supports automatic language detection across 99 languages, selectable Whisper model sizes from Tiny to Turbo, locally cached models, waveform-synchronized timestamps, and in-browser transcript editing.

Limitations and notes

The EarScribe MCP MCP server does not expose a transcription tool, audio upload operation, transcript editing command, or export command. Users must use the EarScribe website for those functions. The local processing and privacy characteristics described in the project apply to the EarScribe browser application: audio stays on the user’s device, and the model is cached locally after its initial load.

Because the MCP component is read-only and provides guidance and links, it fits informational agent workflows rather than automated audio-processing pipelines. No environment variables are documented as required.