jackai-stt-mcp

Transcribe audio by asking for it in plain language:

You: transcribe ~/Downloads/voice.opus

Claude: تمام تمام، موضوع اللغة العربية إن شاء الله محلول…

An MCP server that runs on your own machine, so the assistant can open the file directly. Nothing to run, no commands to memorise — you mention a file and the assistant does the rest.

No uploading, no copying, no base64. Your OpenAI key never leaves your computer.

Arabic works well, including dialect.

Install

Add this to your MCP client's config. uvx fetches and runs the package on first use — nothing to install by hand.

{
  "servers": {
    "jackai-stt": {
      "command": "uvx",
      "args": ["jackai-stt-mcp"],
      "env": {
        "OPENAI_API_KEY": "sk-proj-..."
      }
    }
  }
}

Where the config lives:

Client	File
VS Code	~/Library/Application Support/Code/User/mcp.json (macOS)
Claude Desktop	~/Library/Application Support/Claude/claude_desktop_config.json
Claude Code	.mcp.json in the project, or claude mcp add
Cursor	~/.cursor/mcp.json

Restart the client, then ask it to transcribe something.

Get an API key at platform.openai.com/api-keys. The account needs credit; transcription runs about $0.0045 per minute.

Usage

Nothing to run, no syntax to learn. Talk to the assistant the way you normally would and mention the file:

transcribe ~/Downloads/voice.opus

what does the voice note on my desktop say?

read ./recordings/call.m4a and summarise what the client is asking for

transcribe meeting.mp3 and tell me who said what

The assistant recognises that it needs this tool and fills in the arguments from what you said. You never call it yourself or write out its parameters.

That last example needs speaker labels, which only one model produces. Mention it in passing and the assistant picks it:

transcribe meeting.mp3 with the diarize model

Same for a language it keeps mishearing ("it's Egyptian Arabic"), or names it should spell correctly ("the speakers are Ahmad and Sara"). Ordinary words — the table below is just what those words map onto.

What the assistant fills in

You don't set these by hand; this is here so you know what it can control.

Argument	Default	Notes
file_path	—	Path on this machine. Absolute, relative, or ~/....
audio_url	—	Public http(s) URL, downloaded then transcribed.
audio_base64	—	For short clips. Prefer file_path.
model	gpt-transcribe	See the table below.
language	auto-detect	ISO-639-1 (ar, en). Set only if detection is wrong.
prompt	—	Names or jargon likely in the audio, to steer spelling.

Exactly one audio source per request.

Models

Model	Cost/min	Speaker labels
gpt-transcribe (default)	$0.0045	no
gpt-4o-mini-transcribe	$0.003	no
gpt-4o-transcribe	$0.006	no
gpt-4o-transcribe-diarize	$0.006	yes
whisper-1	$0.006	no

gpt-transcribe is OpenAI's recommended model: cheaper than whisper-1 and more accurate. There is no reason to pick whisper-1 unless you need it specifically.

Formats

flac m4a mp3 mp4 mpeg mpga oga ogg wav webm — up to 25 MB.

.opus files work too. OpenAI rejects that extension even though the bytes are what it accepts as .ogg, so this server renames it in flight. WhatsApp voice notes are all .opus, which is exactly the case that would otherwise fail.

Why it runs locally

A remote MCP server cannot read a file you attached in chat. There is no mechanism in the protocol that carries attachments to a server — it has been an open issue since early 2025, and the proposals to add one are still drafts. Base64 through a tool argument works in theory but dies on client argument-size limits after a second or two of audio.

Running on your machine avoids all of it. The server is a process you own, reading a file you own, with a key you hold.

Development

git clone https://github.com/jack-ai-net/stt-mcp
cd stt-mcp
python3 -m venv .venv && .venv/bin/pip install -e ".[dev]"
.venv/bin/python -m pytest

Tests stub the OpenAI call, so they need no API key and cost nothing.

License

MIT — see LICENSE.

Built by JackAI.