m4x — local, private audio transcription

Turn voice memos and podcasts into text entirely on your own machine. m4x is a thin, fast wrapper around whisper.cpp that ships as both a command-line tool and an MCP server for AI agents.

Think of it as a tiny, local-first alternative to cloud transcription products (Plaud, Otter, etc.) — with one big difference: your audio never leaves your computer. No account, no upload, no API key, no per-minute fee. It runs beautifully on Apple-silicon Macs (Metal-accelerated) but works anywhere whisper.cpp does.

⚡ On an Apple M-series machine, the large-v3-turbo model transcribes roughly an hour of audio in a couple of minutes, fully offline.

Why local?

	m4x (local)	Typical cloud transcriber
Where your audio goes	Stays on your machine	Uploaded to a third party
Account / API key	None	Required
Cost	Free	Per-minute or subscription
Works offline	Yes	No
Summaries	Bring your own LLM (see below)	Built-in (cloud)

Requirements

• Python 3.10+
• whisper.cpp (whisper-cli) — install via Homebrew: brew install whisper-cpp
• A ggml model — download once (see below)

Install

# 1. the transcription engine
brew install whisper-cpp

# 2. a model (large-v3-turbo: great quality/speed; ~1.5 GB)
mkdir -p ~/whisper-models
curl -L -o ~/whisper-models/ggml-large-v3-turbo.bin \
  https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo.bin

# 3. m4x itself
pipx install .        # or: pip install .

CLI usage

# a single voice memo
m4x transcribe ~/Downloads/memo.m4a
# -> writes ~/Downloads/memo.txt

# a whole folder of podcasts, German
m4x transcribe ~/Podcasts --lang de

# print to stdout and pipe straight into your own LLM for a summary
m4x transcribe interview.mp3 --stdout | llm "summarise this into 5 bullet points and action items"

Options: --lang (e.g. en, de, or auto), --model <path>, --out <dir>, --stdout.

MCP usage (Claude Desktop & other agents)

m4x also runs as an MCP server so an agent can transcribe for you. Long files run as a background job — start, then poll.

Add to your MCP client config (e.g. Claude Desktop claude_desktop_config.json):

{
  "mcpServers": {
    "m4x": {
      "command": "m4x-mcp"
    }
  }
}

Tools exposed: transcribe_start(file_path, language) → job_id, transcribe_result(job_id), transcribe_list_jobs().

"Mini-Plaud" pattern: transcribe → summarise

m4x deliberately does transcription only and stays out of the AI-summary business, so you keep full control (and privacy) over that step. Pair it with any LLM you like:

m4x transcribe standup.m4a --stdout | llm "Extract decisions, owners, and due dates as a table"

Or, in an agent: call transcribe_start, poll transcribe_result, then ask the model to summarise the returned text.

Configuration

Everything has sensible defaults; override via environment variables:

Variable	Default	Meaning
M4X_WHISPER_BIN	whisper-cli on PATH, else /opt/homebrew/bin/whisper-cli	Path to the whisper.cpp binary
M4X_MODEL	~/whisper-models/ggml-large-v3-turbo.bin	Path to the ggml model

How it works

m4x shells out to whisper-cli with your chosen model and language, writes a .txt transcript next to the audio (or to --out), and returns the text. The CLI and the MCP server share one small core module — no duplicated logic. That's the whole thing: ~200 lines, no magic.

License

MIT — see LICENSE. Contributions welcome; see CONTRIBUTING.md and SECURITY.md.