Local Whisper

What Local Whisper does

The Local Whisper MCP server gives Claude a local way to convert recordings into text. It accepts WAV, Ogg Opus, and MP3 files without additional software, and can handle M4A and FLAC when ffmpeg is available. With ffmpeg installed, it can also process MP4 and WebM videos by reading their audio tracks without requiring separate extraction.

Transcription results can be returned as plain text, SubRip subtitles (SRT), or timestamped JSON segments. Voice activity detection can skip silence in clean recordings, while the server can transcribe the full input when background noise may conceal quiet speech. This makes the Local Whisper MCP server suitable for voice notes, recorded meetings, classes, and subtitle generation.

How it works

The server runs Whisper locally and downloads the selected model the first time it is needed. It verifies the model against the hash published by Hugging Face. The default model is large-v3-turbo-q8_0, though another model from whisper.cpp's model list can be selected.

GPU execution is available through Vulkan on Windows and Linux and through Metal on Apple Silicon. The server falls back to the CPU when the selected runtime cannot load or GPU use is disabled. After a configurable idle period, it releases the model from memory, including GPU memory. The default idle period is five minutes.

Claude Desktop users can install a platform-specific extension. Claude Code users register the downloaded executable with claude mcp add. A separate optional skill helps Claude recognize when to use the server and how to handle attachments. For claude.ai chats running through the desktop app, attached audio is copied to the server's inbox and removed after transcription.

Setup and configuration

The project distributes native executables, so Python and .NET are not required. Downloads are provided for Windows x64, Intel and Apple Silicon macOS, and Linux x64 and arm64. Linux does not have a Claude Desktop extension in the listed downloads.

Claude Desktop exposes settings for folders, model, and language in the extension settings. Other MCP clients can use environment variables, including LOCAL_WHISPER_MODEL, LOCAL_WHISPER_LANGUAGE, LOCAL_WHISPER_USE_GPU, LOCAL_WHISPER_RUNTIME, LOCAL_WHISPER_IDLE_MINUTES, LOCAL_WHISPER_ALLOWED_ROOTS, LOCAL_WHISPER_INBOX_DIR, and LOCAL_WHISPER_MODELS_DIR. Allowed roots restrict which folders the server can read. The inbox is used for files copied from clients and is cleaned after processing.

Tools and capabilities

• transcribe processes an absolute file path and accepts model, language, output format, and voice-activity-detection options.
• status reports the active device, installed models and sizes, download activity, inbox, allowed folders, and log folder.
• delete_models removes downloaded models to recover disk space; its tool description tells Claude to request confirmation first.
• The Local Whisper MCP server can detect the language automatically or use a specified language code such as en or es.
• SHA-256 values for release downloads are available in SHA256SUMS.txt.

Limitations and notes

Local Whisper requires the target recording to be accessible to the machine running the server. Video support for MP4 and WebM, and audio support for M4A and FLAC, depend on ffmpeg being on PATH. GPU support depends on the operating system and available runtime; CPU fallback is built in.

For claude.ai chats, the computer must be on and the Claude desktop app must be open. A chat that is not running on the computer cannot access the local server. The optional skill helps prevent such chats from attempting cloud transcription instead. Large video attachments may be better referenced by a local path so they do not travel through the chat and back.