MCP SERVERRun whisper.cpp transcription on Windows with GPU acceleration — single files or an automated folder batch, with subtitles, model management and a persistent model server.
LABSCO SUMMARY
Transcribing fifty files one process at a time spends the majority of its runtime loading weights, and a persistent server removes that entirely. Pairing it with a batch mode that advances itself means a folder can be left to run. The setup and hardware checks earn their place for the same reason the batch does: whisper.cpp on Windows with Vulkan has several ways to be quietly not accelerated, and a transcription that is ten times slower than it should be looks identical to one that is working. Everything runs locally, so audio never leaves the machine — which for recordings of meetings or interviews is usually the whole reason to pick this over a hosted service.
A local Whisper transcription client for Windows built around whisper.cpp: single and batch transcription, subtitle generation, model management, and hardware verification.
Windows with whisper.cpp, FFmpeg and a Whisper model — WHISPER_CLI_PATH and WHISPER_MODEL point at them, and check_config verifies all three.
One command — npm install -g whisper-windows-mcp
USE CASES
Verify the CLI, model and FFmpeg are present, and confirm the GPU and Vulkan acceleration are detected rather than silently absent.
Start it as a background job and poll its progress, rather than blocking for the length of the audio.
Start the automated sequential batch over everything untranscribed and let the progress check advance it, with the model server holding the weights in memory throughout.
Generate subtitle files with the language set, rather than a plain transcript you then have to time.
See what is installed with sizes, download another from Hugging Face, and switch the active one without restarting.
Counted once with only startup setup. Other versions, modes or optional credentials may expose a different list.
analyze_mediaAnalyze one or more media files using FFprobe before transcribing. For a single file: returns duration, size, codec, and estimated transcription time on CPU and GPU. For a folder: scans all supported media files and returns a sorted table with the same info for each. Use this to plan batch work, estimate how long transcription will take, or check what's already been transcribed.
check_batch_progressCheck the status of a batch started with start_batch. Automatically advances to the next file when the current one finishes. Returns overall progress, current file, failed files, and elapsed time. Call repeatedly until the batch shows as complete.
check_configVerify whisper-cli.exe, model, and FFmpeg are all available. Run this first if anything fails.
check_progressCheck the status of a background transcription job started with transcribe_audio (background=true). Returns current progress, elapsed time, last processed timestamp, and the transcript when complete. Call this repeatedly until the job shows as complete or failed. ⚠️ Privacy: transcript text returned on completion is processed by Claude's API. Pass privacy_mode=true to return metadata only for this check, regardless of how the job was started.
check_systemDetect GPU hardware and verify Vulkan acceleration is available. Reports GPU name, VRAM, whether the Vulkan binary is installed, and recommends the best Whisper model for your hardware.
download_modelDownload a Whisper model from Hugging Face directly into your models directory. Accepts a model name (e.g. large-v3-turbo, medium.en-q5_0) and handles the download automatically. Downloads only from trusted Hugging Face namespaces (ggerganov/whisper.cpp and ggml-org). After downloading, use switch_model to activate it for the current session.
generate_subtitlesGenerate subtitle files for an audio or video file using whisper.cpp. Set language='auto' to detect the spoken language automatically. Set translate_to_english=true to also generate an English translation subtitle file. Supports SRT and WebVTT (VTT) output formats. When both native and translation are requested, two files are saved: one in the original language and one English translation. Load SRT in VLC via Subtitle → Add Subtitle File. VTT works in web players and HTML5 video. Supports all standard formats plus .3gp and .ts.
list_modelsList all Whisper model files installed in your models directory. Shows filename, size, whether it is currently active, quantization status, and recommended use case for each model. No network calls — reads local filesystem only.
start_batchStart an automated sequential batch transcription of all untranscribed files in a folder. Scans for files without a matching .txt, sorts by duration (shortest first), and processes them one at a time as background jobs. Each file is validated after completion — empty or suspiciously short outputs are flagged. Batch self-advances without polling when each file finishes. Returns a batch ID to use with check_batch_progress. ⚠️ Privacy: when privacy_mode is active, one confirmation is required before the batch starts. All files then process unattended. No transcript text is returned to the API.
switch_modelSwitch the active Whisper model for the current session without restarting Claude Desktop. Accepts a model filename (e.g. ggml-large-v3-turbo.bin) or full path. The model must already be installed in your models directory. Change is session-scoped — does not persist after Claude Desktop restarts.
transcribe_audioTranscribe a single audio or video file using whisper.cpp on Windows. Natively supports mp3 and wav. Automatically converts mp4, mkv, avi, mov, webm, m4a, flac, ogg etc. via FFmpeg — no manual conversion needed. Output defaults to timestamps format (with time codes). For files that may take more than 4 minutes, set background=true to run as a detached job and use check_progress to monitor it. ⚠️ Privacy: transcript text returned by this tool is processed by Claude's API. Pass privacy_mode=true to this tool to enable metadata-only responses per call — no transcript text will be transmitted. Set WHISPER_PRIVACY_MODE=true in env to enable globally. When privacy mode is active, a confirmation is required before every operation.
transcribe_batchTranscribe multiple audio/video files in a folder interactively, one file at a time. Shows a preview of each transcript and waits for confirmation before continuing. Saves each transcript as a .txt file next to its source. Files already transcribed (with matching .txt) are shown as done and skipped. Supported formats: mp3, wav, mp4, mkv, avi, mov, webm, m4a, flac, ogg. NOTE: For large unattended batch jobs, use start_batch instead. ⚠️ Privacy: transcript previews are processed by Claude's API. Pass privacy_mode=true to suppress previews and return metadata only. When privacy mode is active, confirmation is required before each file.
whisper_serverStart, stop, or check the persistent whisper model server. When running, the active model stays resident in VRAM and every transcribe_audio / transcribe_batch call is served over localhost without reloading it — eliminating the per-file model-load cost (a large speedup for many short files). ⚠️ The resident model holds GPU VRAM for the server's entire lifetime, so start it deliberately, do your work, then stop it to hand the GPU back to other applications. While it is running, background jobs, start_batch, generate_subtitles, and lrc/csv or advanced per-call options are refused (they need the one-shot CLI and would contend for the GPU) — stop the server to use those. Bound to localhost only.
BEFORE YOU INSTALL
13 tools when we connected.
Node.js 18 or later
The documented client runs the server on your machine.
CONTEXT COST
Rated by the size of the 13 tool definitions that ride along on every turn, not by how many there are
INSTALLATION
npm install -g whisper-windows-mcp{
"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"
}
}
}
}MCP servers
MuhammadTayyabIlyas
mocomoco-inc
samson-art
AudiopodAI