groq-whisper-mcp

MCP server for audio/video transcription via Groq-hosted Whisper. Word-level timestamps, automatic caching, cost estimation.

9x cheaper than OpenAI Whisper — same model, different API.

Prerequisites

• Python 3.10+

• ffmpeg (for audio extraction from video)

• Groq API key

Related MCP server: Whisper Speech Recognition MCP Server

Installation

git clone https://github.com/bis-code/groq-whisper-mcp.git
cd groq-whisper-mcp
python3 -m venv venv
venv/bin/pip install mcp openai

Claude Code Configuration

Add to your project's .mcp.json:

{
  "mcpServers": {
    "whisper": {
      "command": "/path/to/groq-whisper-mcp/venv/bin/python",
      "args": ["-m", "server"],
      "cwd": "/path/to/groq-whisper-mcp/src",
      "env": {
        "GROQ_API_KEY": "your-groq-api-key"
      }
    }
  }
}

Tools

transcribe_video

Transcribe a video file with word-level timestamps.

Returns full text, word-level timestamps [{word, start, end}], and duration. Results are cached per-project.

estimate_transcription_cost

Estimate cost before transcribing.

Returns duration, estimated cost, and comparison with OpenAI pricing.

How It Works

1. Extracts audio from video via ffmpeg (128kbps MP3, falls back to 64kbps if >25MB)

2. Sends to Groq's OpenAI-compatible API (same openai SDK, different base_url)

3. Returns word-level timestamps with ~0.1s precision

4. Caches results as whisper_words.json alongside the video

Running Tests

PYTHONPATH=src venv/bin/pytest tests/ -v

License

MIT