Hex is a local-first voice dictation app for Apple silicon Macs that transcribes what you say on the machine itself and pastes the result into whatever window you were already typing in. It is for people who dictate a lot and want the words to land in place rather than in a separate transcription window: writers, support and sales people answering long messages, and developers who would rather speak a commit message than type it. It also suits anyone who has a policy, or just a preference, against uploading recordings of their own voice.
What it does
You pick a shortcut, hold it, speak, and release. Hex transcribes locally and pastes the text at the cursor, so it works in the app you are already using rather than only in a supported list. The video walkthrough uses Option: hold Option for a sentence, release, and the text appears; double-tap Option and recording stays on until you tap again, which is the mode for a long paragraph.
Beyond the shortcut, the README names a few things you can set:
- the shortcut itself, the dictation language, and which speech model to use
- Modes, which apply corrections (names and spellings fixed before the text ever lands), application-specific behavior, and optional text processing
- a TypeScript SDK that ships in the repository under sdk/typescript, for driving Hex from your own code
The macOS app is the finished one. The README also lists a Windows beta for x64 and ARM64 PCs and a Linux beta for x86_64 machines running X11 or a compatible wlroots Wayland compositor.
How it works
Hex is a native Rust application, and speech recognition runs on your device: you download a local speech model during setup, and the README is explicit that ordinary dictation needs neither a cloud transcription service nor OpenCode. The README does point to a "Privacy and local data" section covering what stays on the device and what optional features send to a provider, so if you switch on the optional text processing, read that section before assuming everything stays local.
Releases come in two channels. Stable is the current release; Dev is the newest build of the dev branch, published on every change, and it updates itself from the Dev channel.
Nothing in the material gives an accuracy or latency number, and the channel does not run the tools it covers, so the only evidence that it feels instant is the project's own demo.
Getting started
Requirements come straight from the README: an Apple silicon Mac running macOS 15 or newer for the supported app. Windows x64 and ARM64, plus Linux x86_64 on X11 or a compatible wlroots Wayland compositor, exist as betas with their own notes in docs/windows.md and docs/linux.md. The project lists no memory, disk or model-size figures, so leave room for the speech model and expect the first run to need a network connection to fetch it.
Installing is not a clone. The README links to anoma.ly/hex for downloads: a DMG for Apple silicon macOS and installers for Windows x64 and ARM64, each in a Stable and a Dev variant. The source, the contributing guide and the platform docs sit at github.com/anomalyco/hex under the MIT license.
On money: the app is free and open source, and dictation runs on a model on your own hardware, so there is no key to buy for the core feature. The optional text processing sends text to a provider, and if that provider charges, you pay it. The material names no prices.
When to use it / when not
The catch is hardware. This is an Apple silicon Mac tool first: macOS 15 or newer, no Intel path named anywhere in the material, and the Windows and Linux builds are betas, with Linux further limited to x86_64 and to X11 or wlroots Wayland (GNOME's Wayland session is not on that list, so read docs/linux.md before you plan around it).
Use it instead of the Dictation already built into macOS when you want to choose the shortcut, the language and the model yourself, and when Modes earn their keep because you keep dictating the same product names, people and spellings that a generic recognizer mangles.
Don't use it if you are on an Intel Mac, if your desktop is GNOME on Wayland, or if you need one dictation tool to roll out across a mixed fleet today, because the non-Mac builds are not finished.
Alternatives
Apple's built-in Dictation is already on the Mac and costs nothing to try, and for occasional short messages it may be enough; Hex's argument against it is the configurable shortcut, the choice of model, and Modes. Cloud dictation services do the same job on someone else's hardware, which is precisely the trade Hex is built to avoid. If you are comfortable with a terminal, a local speech-to-text command-line tool transcribes files just as well, but you would be building the hold-a-key-and-paste layer yourself.
Take this one seriously if you type on an Apple silicon Mac, dictate often enough to notice the friction, and care where your audio goes: the repository is MIT, actively pushed, and the core loop it promises (hold, speak, paste) is small enough to judge in a day of real use. If you are on Windows, Linux, or an Intel Mac, bookmark it instead and wait for the betas to settle.