TranscribeX MCP Server: Transcribe with Claude and AI Agents

TranscribeX includes an MCP (Model Context Protocol) server. Once it’s connected, an AI assistant such as Claude Desktop or Claude Code can list your models and transcribe audio and video files on your Mac. Transcription runs locally with your TranscribeX models; the assistant receives the transcript text.

Before you start

Open TranscribeX at least once, and download the models you want to use.
You don’t need to turn on the command line. The MCP server runs the tool inside the app directly.

Get the config

Open Settings › MCP. The Client Config box shows the configuration for your copy of the app. Click the copy button next to it. If TranscribeX is in your Applications folder, it looks like this:

{
  "mcpServers": {
    "transcribex": {
      "command": "/Applications/TranscribeX.app/Contents/Helpers/transcribex",
      "args": ["mcp"]
    }
  }
}

If you keep the app somewhere else, use the path shown in Settings.

Claude Desktop

Quit Claude Desktop.
Open (or create) ~/Library/Application Support/Claude/claude_desktop_config.json in a text editor.
Paste the config. If the file already has an "mcpServers" section, add only the "transcribex" entry inside it.
Save the file and open Claude Desktop. TranscribeX’s tools appear in Claude’s tools list.

Claude Code

Run this in Terminal:

claude mcp add transcribex -- /Applications/TranscribeX.app/Contents/Helpers/transcribex mcp

Start a new Claude Code session to use it.

Other MCP clients

Any client that supports local (stdio) MCP servers works. Use the command /Applications/TranscribeX.app/Contents/Helpers/transcribex with the argument mcp.

Tools

transcribex_list_models

Lists your available transcription models and the languages each supports. It takes no parameters. The output is the same as tx models.

transcribex_transcribe_file

Transcribes one audio or video file and returns the result.

Parameter	Type	Default	Notes
`source_file_path`	string	Required	Absolute path to the file.
`model`	string	App default	A model id from `transcribex_list_models`. Leaving it out is recommended.
`language`	string	App default	A language id supported by the model. Leaving it out is recommended.
`should_diarize`	boolean	false	Adds speaker labels. Needs Pro.
`export_format`	string	text	`text`, `markdown`, `json`, `csv`, `pdf`, `srt` or `vtt`. `pdf` requires `export_file`.
`export_file`	string	None	Writes the transcript to this path instead of returning it inline.
`display_mode`	string	text	`text` or `segments`.
`show_timestamp`	boolean	true	Applies to `segments`.
`show_millisecond`	boolean	false	Applies to `segments`.
`compute_units`	string	App setting	`neural`, `gpu`, `cpu` or `all`, for Whisper and Parakeet. Try `gpu` if a model crashes while loading.

The result has the same fields as the tx transcribe output. See command line.

Example prompts

“Transcribe ~/Downloads/interview.m4a and summarize the main points.”
“Which TranscribeX models do I have, and which ones support German?”
“Transcribe ~/Desktop/meeting.mp4 with speaker labels and list the action items for each person.”
“Make SRT subtitles for ~/Movies/talk.mov and save them next to the video.”
“Transcribe every .mp3 in ~/Podcasts/episode-12 and write a show-notes draft.”

Notes

Jobs aren’t kept. Files transcribed through MCP don’t appear in your TranscribeX library. Ask the assistant to save an export file if you want a copy.
Pro models need Pro. If Pro isn’t active, the tool returns “Model is not available for the current license”. Free models work without Pro.
Crashes while loading a model: ask the assistant to retry with compute_units set to gpu.
Long files take time. The assistant waits until transcription finishes. A faster model, such as Parakeet, helps.
Privacy: transcription stays on your Mac, but the transcript returned to the assistant is handled by that assistant’s service.