macwhisper-mcp
Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@macwhisper-mcptranscribe episode.mp3 in English"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
macwhisper-mcp
MacWhisper transcription as an MCP tool, over streamable HTTP.
MacWhisper's mw CLI is a macOS binary that talks to the running app over a local socket. A Linux container can neither execute it nor reach that socket, so a containerised agent cannot transcribe through MacWhisper directly. This server runs on the Mac, exposes one transcribe tool, and listens on loopback. Containers reach it at http://host.docker.internal:10256/mcp.
Transcription runs on the Mac's GPU with whatever model MacWhisper has selected. Nothing is uploaded, and no API key is involved.
Requirements
macOS with MacWhisper installed and licensed
MacWhisper's CLI enabled: Settings → Advanced → Command-Line Tool → Install (puts
mwat/usr/local/bin/mw)
Related MCP server: FunASR MCP Server
Install
bun install
sh install.shinstall.sh generates a launchd user agent that keeps the server running, and a 0600 .token on first run. It prints the wiring command for the token.
The tool
transcribe({ file, language?, model? })file— a file name, not a path, e.g."episode.mp3". Only files sitting directly in the transcribe directory are visible, and symlinks out of it are refused. That keeps this a transcription service rather than a container-to-host file reader.language— ISO 639-1 ("en","zh") or"auto". Defaults to MacWhisper's setting.model—engine:model-idform. Defaults to the model selected in MacWhisper. Worth setting explicitly for non-English audio: Parakeet is English-only, whilewhisper-cpp:ggml-model-whisper-turbohandles both. Runmw modelsfor what's available.
The call blocks until MacWhisper finishes and returns the transcript as text, so a caller has nothing to poll for. A long recording takes several minutes.
Configuration
Variable | Default |
|
|
|
|
| contents of |
The caller puts media in that directory, then names the file. Mount it into the container if the agent does the downloading.
Access
Loopback is not an access boundary — every local process can reach the port — so requests need Authorization: Bearer <token>, compared with timingSafeEqual. With NanoClaw:
ncl groups config add-mcp-server --id <group-id> --name macwhisper \
--url http://host.docker.internal:10256/mcp \
--headers "{\"Authorization\":\"Bearer $(cat .token)\"}"If no token is configured the server accepts every request rather than pretending to be gated.
Check
sh check.shRuns the guards against a live server: no token is refused, path separators are refused, a missing file is named, and a symlink out of the directory is refused.
Licence
MIT
This server cannot be installed
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Related MCP Connectors
Exposes FEDLIN's public security scanners as agent-callable tools over Streamable HTTP.
OCR, transcription, file extraction, and image generation for AI agents via MCP.
Transcribe audio & video to text for AI agents: 100+ languages, speaker labels, webhooks.
Transcode and host video from one prompt; get a playable link back. Agent-native, over MCP.
Related MCP Servers
- AlicenseAqualityCmaintenanceLocal, private audio transcription MCP server enabling AI agents to transcribe audio files entirely on-device without uploading data.3MIT

FunASR MCP Serverofficial
AlicenseNot gradedqualityAmaintenanceLocal speech recognition MCP server powered by FunASR and SenseVoice. It provides privacy-friendly audio transcription through a transcribe_audio tool.20,261MIT- AlicenseNot gradedqualityCmaintenanceMCP server that transcribes local media files and URLs via Deepgram, writing Markdown transcripts to disk with strict per-job cost ceilings and caching. It also provides a CLI and optional transcript formatting through a language model.MIT
- FlicenseNot gradedqualityCmaintenanceExposes local Python tools via a persistent Streamable HTTP endpoint, allowing an OpenAI Agents SDK notebook to discover and invoke them while using Groq for reasoning and tool selection.-