Cloud ASR MCP

A Model Context Protocol (MCP) server for audio transcription using multimodal LLMs—not traditional ASR.

How It Works

Unlike conventional speech-to-text (Whisper, etc.), this MCP uses audio-capable multimodal models (Gemini, GPT-4o Audio, Voxtral) that process audio in a single pass. The key advantage: you can provide text prompt guidance to clean up transcripts on the fly—removing filler words, formatting speaker turns, or applying custom instructions—all in one API call.

Tested: Successfully transcribed 50-minute audio files using Gemini in one pass—no chunking required.

Features

• Single-pass transcription - Multimodal LLMs process audio directly, no chunking needed
• Prompt-guided cleanup - Remove filler words, format output, or apply custom instructions
• Multiple backends - Gemini 2.5, GPT-4o Audio, Voxtral via OpenRouter or direct APIs
• Long-form support - Validated with 50+ minute audio files
• Remote file support - Works with MetaMCP via SSE transport
• Optional file output - Save transcripts as markdown files

Tools

OpenRouter (Recommended)

Tool	Model	Description
openrouter_transcribe	Any	Unified tool with model selection parameter
openrouter_gemini	Gemini 2.5 Flash	Fast and cost-effective
openrouter_voxtral	Voxtral Mini	Excellent voice transcription
openrouter_gpt4o	GPT-4o Audio	High quality transcription

Direct API Access

Tool	Backend	Description
voxtral_transcribe	Mistral	Direct Voxtral API (mini/small models)
gemini_transcribe	Google	Cleaned transcript with filler words removed
gemini_transcribe_raw	Google	Verbatim transcript
openai_transcribe	OpenAI	Whisper gpt-4o-transcribe
openai_transcribe_economy	OpenAI	Whisper gpt-4o-mini-transcribe

Installation

From npm

npm install -g cloud-asr-mcp

From source

git clone https://github.com/danielrosehill/Cloud-ASR-MCP.git
cd Cloud-ASR-MCP/mcp
npm install
npm run build

Configuration

Environment Variables

# OpenRouter (recommended - single key for all models)
OPENROUTER_API_KEY="your-openrouter-api-key"
OPENROUTER_DEFAULT_MODEL="gemini-flash"  # or gemini-pro, gpt-4o-audio, voxtral-mini, voxtral-small

# Direct API keys (optional)
MISTRAL_API_KEY="your-mistral-api-key"
GEMINI_API_KEY="your-gemini-api-key"
OPENAI_API_KEY="your-openai-api-key"

# Transport (for remote/MetaMCP use)
MCP_TRANSPORT="sse"  # or "stdio" (default)
MCP_PORT="3000"

Claude Code MCP Configuration

Add to ~/.claude/settings.json:

Using npm package

{
  "mcpServers": {
    "cloud-asr": {
      "command": "npx",
      "args": ["-y", "cloud-asr-mcp"],
      "env": {
        "OPENROUTER_API_KEY": "your-openrouter-api-key"
      }
    }
  }
}

SSE Mode (for MetaMCP)

MCP_TRANSPORT=sse MCP_PORT=3000 npx cloud-asr-mcp

Then connect via SSE at http://localhost:3000/sse

Usage Examples

Basic transcription

Use openrouter_gemini to transcribe /path/to/audio.mp3

Choose specific model

Use openrouter_transcribe with model "voxtral-small" on /path/to/audio.mp3

Save transcript to file

Transcribe /path/to/audio.mp3 using openrouter_voxtral and save to /home/user/transcripts/

Remote File Support

For remote clients (MetaMCP), send files as base64:

{
  "file_content": "<base64-encoded-audio>",
  "file_name": "recording.mp3"
}

Supported Audio Formats

MP3, WAV, OGG, FLAC, AAC, AIFF, M4A, WEBM, MPEG

Requirements

• Node.js >= 18.0.0
• ffmpeg (for audio downsampling of large files)

License

MIT

Links

• npm Package
• Report Issues