KERNEL X · v0.0.5 alpha · AGPL-3.0

Think out loud.
It's already typed.

Local, private, free dictation. Hold CtrlShiftX, speak, and your words land at the cursor. Transcribed on-device with Whisper. Your voice never leaves your machine.

Get Scribe View source

open source · local-first · no cloud, no accounts, no telemetry

scribe rec

git commit -m "fix the spectrum scaling bug"

auto-paste in 2s edit⏎

Voice is the natural next step of human-computer interaction. As machines become central to how we live and work, speaking to your machine becomes the more efficient alternative to typing. It shouldn't cost you your privacy to do it.

CtrlShiftX record paste

global hotkey · works in any app · release to paste

// how it works

Five states, one gesture.

Press to record, release to transcribe, then confirm. Or let the auto-paste countdown do it for you.

[01]
idle waiting
[02]

recording capturing
[03]
processing whisper
[04]
review confirm
[05]
paste inserted

// features

Everything a dictation tool should be. Nothing it shouldn't.

Review before paste

Every transcription opens a focused card with a soft countdown. Confirm, edit, or discard.

Auto-paste anywhere

The text drops into the app you were using, right where your cursor is.

Searchable history

Every transcript saved to local SQLite. Findable, editable, deletable.

14-day audio archive

A rolling archive of your recordings, garbage-collected after two weeks.

Tray + global hotkey

Lives in your system tray. Record from anywhere without switching apps.

Upload or record

Drop in an MP3, WAV, or FLAC, or capture a note in-app with the same live spectrum.

// continuous learning

It learns your words.

Corrections you make in the review card feed a personalizer, so frequently-fixed words auto-promote and bias the next Whisper run. On-device LoRA fine-tuning is on the way, so each instance evolves to your voice.

“No other open-source project provides this.”

"rec.t"──▶react

"ess cue ell"──▶SQL

"new pip"──▶NumPy

hotword learning · shipped today

// local-first · privacy by design

Your voice never leaves your machine.

Today's dictation tools route your audio to remote servers, exposing your stream of consciousness to infrastructure you don't control before it reaches your screen. Scribe runs entirely on-device: Whisper, ~140MB, optional CUDA on Linux.

no accounts · no servers · no telemetry · local SQLite, local audio

// how it compares

Open, on-device, and it keeps getting better.

tool	license	processing	platforms	learns
ScribeRust, single binary	AGPL-3.0	on-device	mac · linux · win	yes
Wispr Flowsubscription	closed	remote only	mac · win	no
OpenWisprPython / Electron	MIT	on-device	mac · linux · win	no
Nerd DictationPython / VOSK	GPL	on-device	linux (X11)	no
VocalinuxPython	open	on-device	linux	no

// under the hood

No Electron. One static binary.

Idiomatic Rust, a GPU-accelerated UI, and no runtime to ship.

Rust GPUI · Zed Whisper + Parakeet CPAL + PipeWire rustfft SQLite

macOS (Swift FFI) · Linux (X11 / Wayland + PipeWire) · Windows (in progress)

// the open-source roadmap

From alpha to indispensable.

now
Hotword learning

Corrections in the review card bias the next run. Shipped.
next
Cross-platform injection

Linux (X11 + Wayland), macOS Accessibility, Windows UI Automation.
next
On-device fine-tuning

LoRA on your own recordings, so each instance evolves to your voice.
soon
Meeting transcription

Speaker diarization and auto-labelling, on consumer hardware.
soon
scribe-rs crate + community

A library build for FFI/WASM, published to crates.io. Discord for contributors.

// get scribe · v0.0.5 alpha

Open source. Yours to build.

Clone it, build it, run it. FOSS and developer-first.

install

$ curl -sSL https://kernelx.sh/scribe | sh

clone

$ git clone https://git.anvil.forge.computer/X/scribe

build

$ just install

alpha · macOS + Linux · Windows in progress · scribe.rs

Think out loud.It's already typed.