Privacy

Your voice stays
on your computer.

Kudu was built on one rule: what you say is yours. Speech recognition, cleanup and formatting all run on your own machine, and nothing is ever uploaded.

Cloud servernever
✕

Your computer

Microphone

On only while you hold the key

Memory

Audio is never written to disk

On-device model

Speech to text, on your chip

Your app

Clean text, typed in place

The promise

What Kudu never does

These aren’t settings you have to find. They’re how Kudu is built, and the code is open so anyone can check.

Sends your voice anywhere

Audio and text are processed on your computer. There is no Kudu server to send them to.

Keeps your audio

It’s held in memory while it’s transcribed, then discarded. It’s never written to disk.

Asks for an account

No sign-up, no email, no login. Download it and hold the key.

Tracks you

No analytics, no telemetry, no crash reporting, no advertising IDs.

Records your keystrokes

On Linux it reads the keyboard only to notice its own shortcut. Keys are never stored.

Listens in the background

The microphone opens when you press the key and closes when you let go.

On your disk

What’s stored, and where

Everything Kudu keeps is in one folder you own. Delete it and Kudu forgets everything.

Dictation history

The transcript and final text of each dictation, for the Home page. Kept 90 days by default; choose 30 days, 90 days or forever, or clear it.

Usage totals

Words, dictations and time per day for Insights. No text at all, so they survive when history is cleared.

Dictionary and snippets

Your words and spellings, including the ones Kudu learned from your corrections.

Settings

Your shortcut, styles and preferences, as plain JSON files you can read.

macOS

~/Library/Application Support/Kudu

Windows

%APPDATA%\Kudu

Linux

~/.local/share/kudu

Network

The only time it goes online

Once, to download the model you chose. After that, Kudu works on a plane.

The speech model

Downloaded the first time Kudu starts (about 450 MB on a Mac). It runs on your chip from then on.

Smart formatting, if you want it

On macOS 26 it uses Apple’s on-device model. On Windows and Linux you can download a small local model; it’s optional.

No updates phoning home

Kudu doesn’t check in. New versions are on GitHub when you want them.

Clipboard left alone

On Windows, pasted text is kept out of clipboard history and cloud sync, and your previous clipboard comes back.

Reading the screen

Context, kept in memory

Two features read text from the window you’re in. Both stay on your computer and can be turned off.

Spelling names the way they appear

When you start dictating, Kudu reads the visible names and terms in the window, for at most 150 ms, so it can spell them right. They’re kept in memory and discarded after the dictation.

Learning from your corrections

For up to 60 seconds after a dictation, and only until you switch apps, Kudu watches the field it typed into. If you fix a word, only that word is saved to your dictionary.

Don’t take our word for it.

Kudu is open source under the MIT license. Read the code, build it yourself, or watch its network traffic.

Read the sourceSecurity →