Write at the
speed you talk.

Tap a key twice and talk, tap again to stop. The words land at your cursor the moment you stop, in any app. On your machine, for nothing.

macOS, paste into Terminal

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/ZyxWorks/murmurflow/main/install.sh)"

Windows, paste into PowerShell

irm https://raw.githubusercontent.com/ZyxWorks/murmurflow/main/install.ps1 | iex

Then double-tap Control, say something, and tap once more to stop. That is the whole product.

Open source, MIT licensed. That command runs a script you can read first, and so is everything it installs.

Why bother

150wpm

how fast you speak, without trying

40wpm

how fast an average person types

0€ / mo

MIT licensed, with no paid tier above it

A 200-word message is five minutes of typing. Say it instead and you are done in eighty seconds, hands still on the desk. Ten of those a day is most of an hour back, every day.

Conversational speech runs about 130 to 150 words a minute, and an average typist about 40.

What you are not paying for

No subscriptionCloud dictation bills you every month for as long as you use it. This has no plan to upgrade into.

No accountNothing to sign up for, nothing to log into, nothing to cancel later.

No serverwhisper.cpp transcribes on your own machine. Your voice is never uploaded anywhere.

No dependenciesThe Python standard library, ffmpeg, and the OS. That is the whole supply chain.

The text lands the moment you stop

Tap twice, talk, tap again. The whole transcript pastes at your cursor at once, about two seconds after you stop - there is nothing typed while you talk to land in the wrong place if you move the cursor mid-sentence.

step	seconds
warm transcription, model resident	2.3
first transcription after boot, cold	13.3
microphone open, first ever	9.9
speech lost at the start of a clip	0.04

M4 Pro, macOS 26. One model, large-v3-turbo, answers every clip - on one warm server, so the first dictation is as fast as the hundredth.

What it does not do

No compiled helper. Key state is polled through stdlib ctypes. No Xcode, no signing, no notarization.
No always-on microphone. Nothing listens between sentences, and a clip you forget about closes itself after fifteen seconds of quiet - pasting what you said rather than throwing it away.
No wall of settings. A small waveform while the microphone is open, and a soft pulse while it transcribes. No sounds, no presets.
No LLM on the hot path, unless you ask for one.
No Linux. Wayland exposes no global hotkey API at all, so the answer is honest rather than half working.

Privacy, concretely

Your voice is transcribed by your own machine and the recording is discarded the moment that is done, before anything else happens to it. The microphone closes with the key - and by itself after fifteen seconds of quiet, so it cannot be left open.

The transcript is never inspected, filtered, or written down. It is pasted through your clipboard once, and whatever was there before is put back right after. Nothing here makes a network request except the one-off model download.

The source is public, and so is every claim on this page.

Plug your own agent in

Borrow the key. murmurflow trigger prints the key it is on. murmurflow pause --who "your agent's voice chat" stands it down while your agent owns it, and the pause expires by itself.
Reuse the warm model. While dictation is on, whisper.cpp's own whisper-server stays loaded on 127.0.0.1:8479.
Check against the measurements. voice-contract.json is the measured record, and the speech and gesture files are stdlib-only and yours to vendor.

How, in the README.

Who builds this

ZyxWorks is a product studio and forward deployed engineering practice. MurmurFlow is one of the tools we built for our own work and gave away. The paid work is putting the same engineering inside a company.

Forward deployed engineering We set up one process with you on Zyx and hand it over to your own people after 90 days.

What we do for companies →