whisper-windows-mcp

A Windows-native MCP (Model Context Protocol) server that lets Claude Desktop transcribe audio and video files locally using whisper.cpp — with GPU acceleration, multilingual support, and batch processing. All transcription runs locally — no audio, video, or file paths ever leave your machine.

Why does this exist? The popular whisper-mcp package was built for macOS and assumes a Unix environment. It does not work on Windows. This package was written specifically for Windows users who want local AI transcription integrated with Claude Desktop.

What you can do with it

Once installed, you can say things like this directly in Claude Desktop:

• "Transcribe C:\Users\Me\Downloads\meeting.mp3"

• "Transcribe this folder of recordings and save each as a text file"

• "Generate Japanese and English subtitles for this video"

• "Start a batch transcription of everything in this folder"

• "How long will it take to transcribe these files?"

• "Check if GPU acceleration is working"

Related MCP server: Whisper Speech Recognition MCP Server

Requirements

1. Node.js 18 or later — nodejs.org

2. whisper.cpp binaries with Vulkan GPU support — see Step 1

3. A Whisper model file — see Step 2

4. FFmpeg — required for video files and non-WAV/MP3 audio

Step 1 — Install whisper.cpp binaries

Option A — Pre-built Vulkan release (recommended)

Download whisper-vulkan-win-x64.zip from the releases page.

This is a custom-compiled build with Vulkan GPU acceleration enabled. Works with AMD, NVIDIA, and Intel GPUs — no vendor-specific SDK required.

Extract to C:\whisper\Release\. You should end up with:

C:\whisper\Release\whisper-cli.exe
C:\whisper\Release\ggml-vulkan.dll
C:\whisper\Release\ggml.dll
C:\whisper\Release\ggml-base.dll
C:\whisper\Release\ggml-cpu.dll
C:\whisper\Release\whisper.dll

GPU acceleration is automatic — no additional configuration needed.

Option B — Build from source

Requires: Git, CMake, Visual Studio Build Tools 2022+ with "Desktop development with C++", Vulkan SDK from lunarg.com.

git clone https://github.com/ggml-org/whisper.cpp
cd whisper.cpp
cmake -B build -DGGML_VULKAN=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release --target whisper-cli

Copy the binaries from build\bin\Release\ to C:\whisper\Release\.

Note: The official whisper.cpp Windows releases on GitHub do not include a Vulkan build. You must use the pre-built release above or compile from source with -DGGML_VULKAN=ON.

Step 2 — Download a Whisper model

Use download_model in Claude Desktop to install any of these directly. For English-only use: large-v3-turbo (GPU) or medium.en-q5_0 (CPU) are the best starting points. For multilingual use: large-v3-turbo or large-v3-turbo-q5_0 (CPU). English-only models (*.en.bin) output [FOREIGN] on non-English audio and cannot be used for other languages.

Step 3 — Install FFmpeg

FFmpeg is required for video files and non-native audio formats.

Install via winget:

winget install ffmpeg

Or download from ffmpeg.org and add to your PATH.

Verify:

ffmpeg -version

Step 4 — Install this MCP server

npm install -g whisper-windows-mcp

Step 5 — Configure Claude Desktop

Open Claude Desktop → Settings → Developer → Edit Config.

Add the whisper entry:

{
  "mcpServers": {
    "whisper": {
      "command": "npx",
      "args": ["-y", "whisper-windows-mcp"],
      "env": {
        "WHISPER_CLI_PATH": "C:\\whisper\\Release\\whisper-cli.exe",
        "WHISPER_MODEL": "C:\\whisper\\models\\ggml-medium.en.bin"
      }
    }
  }
}

Config file location: C:\Users\YourName\AppData\Roaming\Claude\claude_desktop_config.json

Use double backslashes in all paths.

Save and fully restart Claude Desktop. You should see whisper listed with a green running badge in Settings → Developer.

Step 6 — Verify your setup

In Claude Desktop, ask:

"Check your whisper config"

Then:

"Check your system hardware"

This confirms your GPU is detected and Vulkan acceleration is active.

Available tools

transcribe_audio

Transcribe a single file. Supports blocking (default) or background mode for long files.

check_progress

Monitor a background transcription job started with transcribe_audio (background=true).

Returns elapsed time, last processed timestamp, percentage, and the full transcript when complete.

start_batch

Automated sequential batch transcription of all untranscribed files in a folder. Sorts by duration (shortest first), processes one at a time as background jobs, validates each output.

check_batch_progress

Monitor a running batch. Automatically advances to the next file when the current one finishes. Returns overall progress, current file with timestamp, ETA, and any failed files.

transcribe_batch (interactive)

Process files one at a time with a preview and confirmation before each. Useful when you want to review as you go.

generate_subtitles

Generate SRT subtitle files. Supports automatic language detection and English translation output.

When both native and translation are requested, two files are saved next to the source:

• filename.ja.srt — original language

• filename.en.srt — English translation

Whisper's built-in translation only translates to English. For other target languages, translate the .srt file contents separately.

analyze_media

Analyze files before committing to transcription. Returns duration, size, codec, and estimated transcription time on CPU and GPU. For folders, shows all files in a sortable table with transcription status.

check_config

Verify whisper-cli.exe, the model file, and FFmpeg are all accessible. Run this first if anything is failing.

list_models

List all Whisper model files installed in your models directory. Shows filename, size, whether it is currently active, quantization status, and recommended use case. No network calls — reads local filesystem only.

download_model

Download a Whisper model directly from Hugging Face into your models directory. Accepts a model name (e.g. large-v3-turbo, medium.en-q5_0) and handles the download automatically. Only downloads from trusted Hugging Face namespaces. After downloading, use switch_model to activate it.

switch_model

Switch the active Whisper model for the current session without restarting Claude Desktop. Change is session-scoped — does not persist after restart. To make permanent, update WHISPER_MODEL in your config.

check_system

Detect GPU hardware and verify Vulkan acceleration is available. Reports GPU name, VRAM, whether ggml-vulkan.dll is present, and recommends the best model size for your hardware.

Supported formats

GPU acceleration

The pre-built Vulkan release enables GPU acceleration automatically. Tested on AMD Radeon RX Vega 56 (GCN 5th gen). Any GPU with Vulkan 1.0+ support should work, including NVIDIA and Intel Arc.

Performance comparison (medium.en model, ~5 minute audio file):

GPU utilization during transcription is typically 15–20%, dropping back to idle between files. CPU stays around 15%.

Multilingual support

Whisper can auto-detect the spoken language and transcribe in that language. The built-in translation model translates to English only.

For best multilingual accuracy, use the large-v3 model. English-specific models (*.en.bin) cannot detect or transcribe other languages.

Example — foreign language video with subtitles:

1. Ask Claude to generate subtitles with language=auto and translate_to_english=true

2. Whisper detects the language and generates a native-language SRT

3. A second pass generates an English translation SRT

4. Load either file in VLC via Subtitle → Add Subtitle File

Designed for free-tier users

This tool is built to minimize Claude API interactions. The entire transcription workflow — scan, analyze, queue, run, validate — is designed to require as few Claude interactions as possible. Heavy lifting is done locally on your machine.

Optional environment variables

Troubleshooting

See TROUBLESHOOTING.md for detailed solutions. See PRIVACY.md for compliance guidance if you handle regulated content.

Quick checklist:

• Paths in config use double backslashes (C:\\whisper\\...)

• whisper-cli.exe exists at the configured path

• Model .bin file exists at the configured path

• FFmpeg is installed and in PATH (ffmpeg -version works)

• Claude Desktop was fully restarted after editing config

• Whisper shows running in Settings → Developer

Security and Privacy

whisper-windows-mcp is designed with security as a core principle.

Audio never leaves your machine. No audio or video files, no file paths, and no telemetry are ever transmitted to any server. No cloud APIs are required for core functionality.

Transcript text and the API boundary. When a tool response includes transcript text, that text is processed by Claude's API — it leaves your local machine. For most users (public content, podcasts, streaming recordings) this is expected behavior. If you handle medical, legal, financial, or other regulated recordings, see PRIVACY.md for compliance guidance and configuration options.

A WHISPER_PRIVACY_MODE environment variable is planned that will restrict all tool responses to metadata only (filename, duration, word count) — no transcript text will be returned to Claude. This is the correct configuration for regulated or confidential content.

Input validation. All file paths are validated before use — UNC paths (\\server\share) and directory traversal sequences (..) are rejected. Files over 10 GB are rejected to prevent resource exhaustion.

Transcript injection awareness. Audio files can contain spoken content that, when transcribed, resembles instructions. Claude's built-in defenses handle this, but it is worth knowing that transcript content is treated as data — never as instructions — by the MCP server itself.

Model downloads are restricted. The download_model tool only downloads from two trusted Hugging Face namespaces (ggerganov/whisper.cpp and ggml-org). Arbitrary URLs are rejected. Redirects are validated against an allowlist before following.

Model switching is sandboxed. switch_model only accepts .bin files within the configured models directory. Paths outside that directory are rejected.

No new network dependencies. Model downloads use Node.js built-in https — no external HTTP libraries are added to the package.

License

Non-commercial use: MIT — free for personal, educational, and non-commercial use. See LICENSE.

Commercial use: A separate commercial license is required for any business, professional, or revenue-generating use. See LICENSE-COMMERCIAL.md for terms and contact information.

Contributing

Pull requests welcome. See ROADMAP.md for planned features.

If you've tested GPU acceleration on hardware not listed above, please open an issue with your results — GPU model, VRAM, model size, and observed throughput.