Whisper Windows MCP README
The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Whisper Windows MCP listing page.
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-mcppackage 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"
- "Transcribe this file in privacy mode"
Requirements
- Node.js 18 or later — nodejs.org
- whisper.cpp binaries with Vulkan GPU support — see Step 1
- A Whisper model file — see Step 2
- 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:
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.
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
| Model | Size | Speed | Accuracy | Best for |
|---|---|---|---|---|
ggml-tiny.en.bin | 75 MB | Very fast | Basic | Quick tests |
ggml-base.en.bin | 142 MB | Fast | Good | Everyday English |
ggml-small.en.bin | 466 MB | Moderate | Better | Important recordings |
ggml-medium.en.bin | 1.5 GB | Fast on GPU | Very good | Best quality English |
ggml-large-v3-turbo.bin | 1.6 GB | Fast on GPU | Excellent | Recommended for English GPU batch work — ~6x faster than large-v3 with minimal accuracy loss |
ggml-large-v3.bin | 2.9 GB | Fast on GPU | Excellent | Multilingual, maximum accuracy |
ggml-medium.en-q5_0.bin | 514 MB | Fast | Very good | Best CPU-only English option — high accuracy at low memory |
ggml-large-v3-turbo-q5_0.bin | 547 MB | Fast | Excellent | Best CPU-only multilingual option |
ggml-large-v3-q5_0.bin | 1.1 GB | Moderate on CPU | Excellent | Multilingual, CPU-friendly |
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:
Or download from ffmpeg.org and add to your PATH.
Verify:
Step 4 — Install this MCP server
Step 5 — Configure Claude Desktop
Open Claude Desktop → Settings → Developer → Edit Config.
Add the whisper entry:
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.
| Parameter | Description |
|---|---|
file_path | Absolute path to the file (required) |
language | Language code (en, ja, es, etc.) or auto to detect. Default: en |
output_format | timestamps (default), text, json, srt, vtt, lrc, or csv |
save_to_file | Save transcript as .txt next to the source file |
background | Run as detached job — returns a job ID immediately. Use check_progress to monitor. Recommended for files over 10 minutes. |
privacy_mode | Override privacy mode for this call. true = metadata only, no transcript text transmitted. false = return text even if WHISPER_PRIVACY_MODE=true globally. Omit to use global setting. |
threads | CPU thread override |
temperature | Sampling temperature 0.0–1.0. Default 0.0 (deterministic). |
prompt | Prior context string — improves accuracy for domain-specific vocabulary or speaker names. Example: "Names: Keemstar, DramaAlert." |
condition_on_prev_text | Re-enable context conditioning between segments. Default false. |
beam_size | Beam search width. Higher = more accurate, slower. Default 5. |
best_of | Candidate sequences evaluated. Default 5. |
gpu_device | GPU device index for multi-GPU systems. Default 0. |
processors | Parallel processor count. Default 1. |
word_timestamps | One word per timestamped segment. Useful for clip alignment. |
max_segment_length | Max segment length in characters. |
diarize | Stereo speaker diarization — requires stereo audio with speakers on separate channels. |
tinydiarize | Mono speaker-turn detection — marks [SPEAKER_TURN] at speaker changes on single-channel audio. Requires a tdrz model: download_model small.en-tdrz, then switch_model ggml-small.en-tdrz.bin. |
vad_model | Path to Silero VAD model .bin. Strips silence before transcription — reduces hallucinations on noisy files. |
offset_t | Start offset in milliseconds. |
duration | Process duration in milliseconds from offset. |
Output formats:
timestamps— timestamped segments, e.g.[00:00:01.230 --> 00:00:04.560] Hello world(default)text— plain text, no time codesjson— structured JSON (blocking mode only)srt— SubRip subtitle file saved next to sourcevtt— WebVTT subtitle file saved next to sourcelrc— LRC lyrics/karaoke format saved next to sourcecsv— CSV with timestamps saved next to source
check_progress
Monitor a background transcription job started with transcribe_audio (background=true).
Returns elapsed time, last processed timestamp, and the full transcript when complete.
| Parameter | Description |
|---|---|
job_id | Job ID returned by transcribe_audio |
privacy_mode | Override privacy mode for this check. true = metadata only, regardless of how the job was started. |
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. Batch self-advances when each file finishes — no polling required.
| Parameter | Description |
|---|---|
folder_path | Path to folder (required) |
language | Language code. Default: en |
threads | CPU thread override |
output_format | timestamps (default) or text |
privacy_mode | Override privacy mode. One confirmation required before batch start; all files then process unattended. No transcript text returned. |
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, and any failed files.
| Parameter | Description |
|---|---|
batch_id | Batch ID returned by start_batch |
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.
| Parameter | Description |
|---|---|
folder_path | Path to folder (required) |
file_index | Which file to process (1-based). Omit to list files first. |
language | Language code. Default: en |
recursive | Include subfolders |
output_format | timestamps (default) or text |
privacy_mode | Override privacy mode. Confirmation required before each file; metadata only returned. |
generate_subtitles
Generate subtitle files. Supports automatic language detection and English translation output. Outputs SRT (widest compatibility) or WebVTT (web and HTML5 video).
| Parameter | Description |
|---|---|
file_path | Path to file (required) |
language | Language code or auto to detect. Default: en |
output_format | srt (default) or vtt |
translate_to_english | Also generate an English translation subtitle file. Only applies when source is not English. |
background | Run as detached background job. Returns a job ID for check_progress. |
threads | CPU thread override |
When both native and translation are requested, two files are saved next to the source:
filename.ja.srt— original languagefilename.en.srt— English translation
Whisper's built-in translation only translates to English. For other target languages, translate the subtitle 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.
| Parameter | Description |
|---|---|
path | Path to a single file or folder (required) |
sort_by | For folders: duration (default), name, or size |
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. Only downloads from trusted Hugging Face namespaces. After downloading, use switch_model to activate it.
| Parameter | Description |
|---|---|
model_name | Model name to download, e.g. large-v3-turbo, large-v3-turbo-q5_0, medium.en-q5_0 |
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.
| Parameter | Description |
|---|---|
model_name | Model filename (e.g. ggml-large-v3-turbo.bin) or full path. Must be a .bin file in the configured models directory. |
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.
whisper_server
Start, stop, or check the persistent model server (whisper.cpp's whisper-server). While running, the active model stays resident in VRAM and every transcribe_audio / transcribe_batch call is served over localhost with no per-file model reload — a large speedup when transcribing many short files, where the one-time model-load cost otherwise dominates.
| Parameter | Description |
|---|---|
action | start — launch with the active model resident; stop — shut down and free VRAM; status — report running state, resident model, port, and uptime. |
- ⚠️ The resident model holds GPU VRAM for the server's whole lifetime. Start it deliberately, do your work, then
stopit to hand the GPU back to other applications sharing the card. Stopping performs a full kill so VRAM is actually released. switch_modelwhile the server is running hot-swaps the resident model in place (no restart).- Bound to
127.0.0.1only — never exposed on the network. - While the server is up, operations that need the one-shot CLI — background jobs,
start_batch,generate_subtitles,lrc/csvoutput, and advanced per-call options the HTTP API doesn't honor (beam_size,best_of,word_timestamps,diarize,tinydiarize,vad_model,offset_t,duration) — are refused with a "stop the server first" message rather than silently ignored, so no second engine ever contends for the GPU. - Requires
whisper-server.exe(ships alongsidewhisper-cli.exe). Configure withWHISPER_SERVER_PATH/WHISPER_SERVER_PORTif needed.
Supported formats
| Type | Formats |
|---|---|
| Native (no conversion) | mp3, wav |
| Video (auto-converted via FFmpeg) | mp4, mkv, avi, mov, webm, flv, wmv, m4v, ts, 3gp |
| Audio (auto-converted via FFmpeg) | m4a, ogg, flac |
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 (large-v3 model, ~14 minute audio file):
| Hardware | Time |
|---|---|
| CPU only (Ryzen 7 2700x, 8 threads) | ~22 minutes (estimated) |
| GPU (Vega 56 via Vulkan) | ~3m 22s |
GPU utilization during transcription is typically 15–20%, dropping back to idle between files.
Supports Windows 10 and Windows 11. No Windows 11-specific configuration is required — the tool makes no Win32 API calls and runs on either OS.
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:
- Ask Claude to generate subtitles with
language=autoandtranslate_to_english=true - Whisper detects the language and generates a native-language SRT or VTT
- A second pass generates an English translation
- Load the SRT in VLC via Subtitle → Add Subtitle File, or use the VTT in any web player
Privacy and compliance
whisper-windows-mcp includes a built-in privacy architecture for sensitive and regulated content.
Audio and video never leave your machine. This guarantee is unconditional.
Transcript text is different — when returned inline in a tool response, it is processed by Claude's API. For most users this is expected behavior. For regulated content (medical, legal, financial, corporate), privacy mode prevents this.
Privacy mode restricts all tool responses to metadata only (filename, word count, save path). No transcript text is transmitted to Claude's API under any circumstances. Enable per-call with privacy_mode=true on any transcription tool, or globally via WHISPER_PRIVACY_MODE=true in your config.
Consent gate — on first use per session in standard mode, a full privacy disclosure is shown before any transcript text is returned. You must explicitly confirm before proceeding. Set WHISPER_CONSENT_ACKNOWLEDGED=true in your config to skip this for non-sensitive content.
See PRIVACY.md for full compliance guidance (HIPAA, GDPR, attorney-client privilege, FERPA, SOX, PCI-DSS).
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
| Variable | Description |
|---|---|
WHISPER_CLI_PATH | Path to whisper-cli.exe (required) |
WHISPER_MODEL | Path to model .bin file (required) |
WHISPER_THREADS | CPU thread count override |
WHISPER_GPU_DEVICE | Vulkan device index to pin transcription to, for multi-GPU systems (the Vulkan enumeration index — check whisper-cli's startup log; not the Windows GPU order). Overridable per-call with gpu_device. See TROUBLESHOOTING.md. |
WHISPER_FOREGROUND_MAX_SEC | Foreground-transcription cutoff in seconds (default 210). Files estimated to run longer are routed to background mode instead of risking Claude Desktop's ~4-minute tool timeout. |
FFMPEG_PATH | Path to ffmpeg if not in system PATH |
WHISPER_SERVER_PATH | Path to whisper-server.exe for the persistent model server (default: alongside whisper-cli.exe). See the whisper_server tool. |
WHISPER_SERVER_PORT | Localhost port for the persistent model server (default 8571). Always bound to 127.0.0.1. |
WHISPER_PRIVACY_MODE | When true, all tool responses return metadata only — no transcript text transmitted to Claude's API. For regulated or confidential content. Can be overridden per-call with the privacy_mode parameter. See PRIVACY.md. |
WHISPER_CONSENT_ACKNOWLEDGED | When true, skips the one-time session consent disclosure shown before transcript text is returned. Set after you understand the privacy boundary and no longer need the reminder. Has no effect when privacy mode is active. |
Security
Binary verification. To verify the integrity of the whisper-cli.exe binary in the pre-built release, check its SHA256 hash in PowerShell:
The expected hash for the v1.4.0 release binary is documented in the releases page.
Input validation. All file and folder paths are validated before use, on every tool that takes one — UNC paths (\\server\share) and directory traversal sequences (..) are rejected. Files over 10 GB are rejected to prevent resource exhaustion. job_id and batch_id are checked against the exact server-minted format before they are used to build any file path, so a crafted ID cannot traverse out of the jobs directory.
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. Because transcribed content can still influence which tools Claude calls next, path/ID validation is applied defensively rather than trusting the single-user assumption alone.
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. (Downloads are not yet verified against a per-model SHA256 digest — see SECURITY.md.)
Model selection is sandboxed. Both switch_model and the transcribe_audio model override only accept .bin files within the configured models directory. Paths outside that directory are rejected via normalized path containment.
No PATH shadowing. System binaries the server invokes on your behalf (tasklist, wmic) are called by absolute System32 path so they can't be shadowed by a same-named executable earlier on PATH.
See SECURITY.md for the full security policy.
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.exeexists at the configured path- Model
.binfile exists at the configured path - FFmpeg is installed and in PATH (
ffmpeg -versionworks) - Claude Desktop was fully restarted after editing config
- Whisper shows running in Settings → Developer
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 COMMERCIAL-LICENSE.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.