Dictation for Linux, in early access

Hold a keyand ramble.It lands written.

Wrenscribe listens while you hold the shortcut, cuts the ums and the “no, wait” corrections, and pastes finished text into whatever app has focus. The speech model and the editor both run on your computer, so your voice never leaves it.

Early access opens on Linux first. No spam, unsubscribe any time.

  • No account
  • No cloud
  • GNOME on Wayland
Claude CodeGhostty, coding agent style
What you saidlistening

um okay so refactor the bridge so every request is uh validated before it hits the handler, actually no, before it reaches the engine

What lands in Claude Code

Refactor the bridge so every request is validated before it reaches the engine.

Pastes into
  • Claude Code
  • Codex
  • Ghostty
  • Ptyxis
  • Firefox
  • Brave
  • Chrome
  • Discord
  • Signal
  • and any other text box

How it works

One shortcut. No window to babysit.

Wrenscribe sits in the background with a small pill at the bottom of the screen. You never switch to it to dictate, and it never takes focus from the app you are typing in.

  1. 1

    Hold the shortcut

    Any two-key combination you pick. A short tap starts hands-free recording instead, a second tap ends it, and Escape throws it away.

    Ctrl + Space
  2. 2

    Talk like you talk

    Parakeet, an open speech model, transcribes on your CPU while you are still talking, a stretch at a time at each pause. A five-minute ramble does not leave you waiting at the end. A sound marks the start and the stop, and the pill shows your levels.

  3. 3

    Let go

    A small language model on your machine edits the transcript. Then it is pasted where your cursor is, and your clipboard is put back the way it was.

Styles

It knows where you are typing.

A prompt for Claude Code should not read like a Discord message. Wrenscribe checks the focused window and writes for it. Any app can be given a style of its own.

You said

um so the flatpak build fails on the ort step, can you check build slash flatpak and uh tell me why

Wrenscribe pasted
> The Flatpak build fails on the ORT step. Can you check build/flatpak and tell me why?

Plain text, no greeting. Paths and identifiers stay exactly as said, and terminals get Ctrl+Shift+V.

Local by design

Your voice stays on this machine.

There is no Wrenscribe server to send it to. Once the models are on disk, a dictation does not touch the internet at all.

Speech
Parakeet TDT 0.6B, on your CPUMeasured at 31 times faster than real time on the machine it was built on.
Cleanup
Gemma 4 E2B through llama.cppOn your graphics card over Vulkan, or the CPU if that is all there is.
Account
There isn't oneNo sign-in, no API key, no monthly word quota.
History
A SQLite file in your home folderRaw and cleaned text for every dictation. Audio is deleted after 7 days unless you change it.
Network
Model downloads, onceThe only thing Wrenscribe fetches. Audio and text never go out.
Telemetry
NoneNo analytics and no crash reports. Logs stay in the app's own folder.

The polished dictation apps send your voice to a data center to tidy it up. Wrenscribe does the tidying on the computer in front of you.

Your words

Names spelled your way.

Speech models trip on product names and jargon. Wrenscribe corrects them after the fact, from words you teach it and a built-in list where every entry was measured against real mishearings.

Dictionary

Add a word once, with the ways it comes out wrong. Sound-alike matching catches the near misses.

  • clod code->Claude Code
  • Sequalite->SQLite
  • D V->DB

Snippets

Say a trigger phrase and the saved text is pasted instead. Good for anything you type every day.

you say: “standup template”Yesterday:
Today:
Blocked on:

Command mode

Select some text, hold the command shortcut and say what to do with it: “make this shorter”, “turn this into bullet points”. The rewrite replaces the selection.

History and replay

Every dictation is kept stage by stage: what the speech model heard and what the cleanup made of it. When one comes out wrong, Report a problem records what you expected, and replay runs it through the pipeline again.

$ wrenscribe replay 52 --json

Linux first

Built on a Linux desktop, for one.

Dictation apps tend to reach Linux last, if they reach it at all. Wrenscribe started on GNOME and Wayland, which is where the hard parts are: global shortcuts, pasting into other apps, and an overlay that never steals focus.

  • Hold-to-talk through GNOME's GlobalShortcuts portal. No input group, no udev rules.
  • A GNOME Shell extension draws the pill, pastes the text and restores your clipboard.
  • Ships as one Flatpak with the extension inside.
  • KDE, macOS and Windows are on the roadmap, in that order of difficulty.
How it works on Linux ->
early access install
# early access builds ship as one Flatpak
$ flatpak install --user wrenscribe.flatpak

# GNOME loads extensions at login, so once:
$ gnome-session-quit --logout

# after that, hold your shortcut and talk

Pricing

$0

Free for everyone in early access, with every feature and no word limit. Your computer does the work, so a dictation costs us nothing. There is no subscription, and none is planned.

How pricing will work ->

Questions

Fair questions.

When can I use it?

Wrenscribe runs every day on the machine it was built on, and early access opens on Linux first. Join the waitlist and you will get one email when there is a build to install.

Which systems does it run on?

Linux with GNOME on Wayland. It is built and tested on Bazzite with GNOME 50. Other Linux desktops, macOS and Windows are planned; the shared core is already split from the Linux parts to make those ports possible.

Does it work offline?

Yes. The first run downloads the speech and cleanup models. After that, recording, transcription, cleanup and pasting all happen on your computer with no network involved.

What hardware does it need?

Speech recognition runs on the CPU. The cleanup model prefers a graphics card through Vulkan and falls back to the CPU, which is slower. Exact requirements will be published with the first public build.

What does the cleanup actually change?

It removes fillers, applies your self-corrections ("Tuesday, no, Wednesday" becomes "Wednesday"), turns spoken lists into real ones, and fixes punctuation and casing for the app you are in. It does not rewrite what you meant. History keeps the raw transcript next to the cleaned text, so you can always check.

Which languages does it understand?

English only for now. Everything, from the speech model to the built-in vocabulary, is tuned for English.

Is it open source?

Not yet. The code is private while the first version settles.

Waitlist

A tiny bird with a very loud song.

That is the whole idea: a small program that does one job very well. Leave your email and you will hear from us once, when there is a build to install.

Early access opens on Linux first. No spam, unsubscribe any time.