How to use Hex for voice dictation

By

Updated

Tagged: AI

Hex is Anomaly's local voice dictation app for Mac, Windows, and Linux. Hold the shortcut, speak, and the transcript pastes into the focused text field.

Hex is a voice dictation app. You hold a shortcut, speak, and release. It transcribes on your computer and pastes the text into the app that has focus.

It comes from Anomaly, the company behind OpenCode. The download page is copyright Anomaly, the GitHub repository is anomalyco/hex, and the AUR package describes it as Hex by Anomaly. The repository is MIT licensed, with Christopher Langton named as the copyright holder.

On October 31, 2026, the OpenCode account announced a stable release, calling Hex fast, local, and available on Mac, Windows, and Linux.

This walkthrough covers Hex 3.1.2, using the download page, the 3.1.2 release, and the docs in that tag.

The download page headline is “Speak. It appears.” Under that, it says local voice dictation for Mac: hold Option, say what you mean, and Hex pastes it where you are. The README describes the same gesture as hold a shortcut, speak, and release.

If you are choosing among Mac dictation apps, the comparison lives in the best AI dictation and voice to text apps for a Mac. Hex is one entry there. This page is the install and daily-use walkthrough.

How a Hex dictation becomes textYou hold the shortcut and speak. After you release, a local speech model transcribes the audio and Hex pastes the text into the focused app.Hold the shortcutSpeakReleaseLocal speech modelPaste into thefocused app

Install Hex

Hex has two channels. Stable is the current release, and dev is the newest build of the dev branch, which updates itself from the dev channel.

The commands below are the stable ones from anoma.ly/hex. The 3.1.2 docs still show installer URLs under https://anoma.ly/hex/files/. Both the site script and that older path install the stable release when the release service has one, and the newest dev build when it doesn’t. Set HEX_INSTALL_CHANNEL=dev to stay on the dev channel, or HEX_INSTALL_CHANNEL=stable to require the stable release.

macOS

Hex for Mac is an Apple silicon app for macOS 15 or later. The download table has no Intel build.

Download the signed DMG from the site, open it, and drag Hex.app into Applications. Or install the same signed app with Homebrew:

brew install --cask anomalyco/tap/hex

Launch Hex from Applications and finish setup. The cask leaves updates to Hex’s own updater. Homebrew does not replace another app that is already named Hex.app.

You can also check by hand with Hex > Check for Updates…. Each GitHub release has the same signed files if a download is blocked. App releases from 3.0.0 onward use v… tags. Through 2.3.1 they used app-v… tags.

If you still have the original Swift app, install this one manually and set it up again. Settings and history are not transferred. Quit the old app before you dictate, so the two shortcuts do not fight.

Windows

The Windows build is a beta for Windows 10 and 11, on x64 and ARM64. An x64 PC needs a 2013 or newer processor with AVX2, which means Intel Haswell, AMD Excavator, or later. On an older processor Hex shows a message and quits.

The download page offers an installer for x64 and one for ARM64. The PowerShell command on that page is:

powershell -ExecutionPolicy Bypass -Command "iwr https://anoma.ly/hex/install-windows.ps1 -OutFile install-windows.ps1; ./install-windows.ps1"

The installer needs no administrator rights. It installs for the current user, adds a Start menu entry, and opens Hex. The first run downloads a speech model.

The installer isn’t code-signed yet, so SmartScreen may say it protected your PC. The Windows guide says to choose More info, then Run anyway.

x64 transcription uses the GPU through Vulkan, so you need a current NVIDIA, AMD, or Intel graphics driver. ARM64 transcribes on the CPU. Neither build needs the Visual C++ redistributable.

Linux

The Linux build is also a beta. It targets x86_64 and ARM64. An x86_64 machine needs the same AVX2 processor requirement as Windows. Published builds are made on Ubuntu 24.04 and need glibc 2.39 or later, which means Ubuntu 24.04, Debian 13, Fedora 40, current Arch, or newer.

The download page’s script installs a user-local copy that updates itself:

curl -fsSL https://anoma.ly/hex/install-linux.sh | sh

Read the script before you pipe it if you want to see what it does. It runs without root and verifies the signed release feed, the file size, and the SHA-256. Then it puts hex on ~/.local/bin, adds a desktop launcher, and can start at login.

The same hex executable is a CLI for the running app. hex status prints hex: Hex is not running; run hex start and exits 3 when the app isn’t running, and the CLI doesn’t open the microphone by itself.

Debian and Ubuntu can install the .deb. Fedora and openSUSE can install the .rpm. The site’s x64 Debian command is:

curl -fLo /tmp/hex.deb https://anoma.ly/hex/download/linux-x64-deb && sudo apt install /tmp/hex.deb

The ARM64 package is linux-arm64-deb. Fedora uses linux-x64-rpm or linux-arm64-rpm, then sudo dnf install on the downloaded file. openSUSE uses zypper install on the same .rpm. There is no apt or dnf repository. When a new stable release exists, Settings shows the command that installs that package. Hex does not replace a package-installed executable by itself.

Arch can use the AUR package. The download page’s command, once the channel is stable, is:

yay -S hex-bin

hex-bin points at the 3.1.2 Linux binaries. Don’t confuse it with the official hex package, which is a hex dump tool and installs hx.

Nix is the other package tab on the download page:

nix profile install github:anomalyco/hex

The Nix guide says the flake packages Hex for x86_64-linux. In-app updates don’t touch a Nix install, so you update the flake input and rebuild. The guide’s Wayland contract doesn’t cover GNOME and KDE, which the Linux guide’s shortcut section handles instead.

The first launch downloads a speech model. You can also run hex model install after the binary is on your path.

Permissions

macOS

Setup asks for Microphone, so Hex can record while you dictate, and Accessibility, so it can see its shortcut in every app and paste into the foreground app. For Accessibility, drag the Hex tile into the list. Hex does not need Input Monitoring.

Annotate needs Screen Recording, and macOS asks for it the first time you start Annotate. Until then nothing is recorded.

A permission change can require quitting Hex and opening it again. The bundle id is anomaly.hex. A Mac that still has the older com.kitlangton.hex2 entries, from releases through 2.3.1, is looking at the previous app. Remove those old rows and allow Microphone and Accessibility again for the new one.

Windows

The Windows guide has no Privacy settings steps. It says the microphone records only while you dictate, and that the shortcut works system-wide when Settings reads Enabled. The SmartScreen warning during install is covered above.

Linux

The Linux guide does not document a system microphone prompt like macOS Accessibility. Audio goes through ALSA, including PipeWire with ALSA compatibility. The window still needs a Vulkan driver even when transcription falls back to the CPU. vulkan-icd-loader alone is not a driver.

Shortcuts are the permission that matters. On X11, Hex grabs the shortcut you record. On Wayland desktops with the GlobalShortcuts portal, including GNOME and KDE, the desktop grants, reassigns, or refuses the shortcut. Settings lists the result under Desktop shortcuts.

Paste on wlroots compositors uses wtype and the virtual keyboard protocol. On KDE and GNOME it uses the RemoteDesktop portal. The first paste there asks you to allow remote control. Install wl-clipboard as well. If paste cannot be delivered, the transcript stays on the clipboard for Ctrl+V.

Do not add your user to the input group to make shortcuts work. That was the 2.3.1 Wayland approach. Current builds use the portal or a keybinding that runs hex.

The hold-to-talk shortcut

The default chord depends on the system:

SystemDefaultWhat hold does
MacOptionHold Option, speak, release. Hex pastes at the focus.
WindowsCtrl+SpaceHold both keys, speak, release Space.
LinuxAlt+SpaceHold Alt+Space, speak, release.
GNOME, new setupsCtrl+Alt+SpaceGNOME keeps Alt+Space for the window menu, so Hex starts here instead.

Escape never cancels on any system, so every dictation finishes and pastes.

On a Mac, double-tap Option to keep recording, then press Option again to finish. A tap under 300 ms records nothing. Option-Shift-V pastes the last completed dictation again.

On Windows, Hold is the mode those keys use out of the box. Toggle is also in Settings: press Ctrl+Space to start and press it again to finish. In Hold mode, tap the shortcut twice within 300 ms to lock recording, then press it again to finish. Paste last is Ctrl+Alt+V. Holding a modifier by itself does nothing. Hex consumes Ctrl+Space while the chord matches, so other apps do not see it. Rebind it if you need that chord elsewhere. Share dictation shortcut with other apps, under Advanced, lets another app see the same key, which can also fire that app’s own binding.

On Linux, double-tap the shortcut to keep recording, then press the same chord again to finish. Ctrl+Alt+V pastes the last dictation. Toggle mode starts and stops with single presses. Double-tap only waits for two complete taps before it records.

Hyprland accepts the portal shortcut without a key, so nothing happens until you bind one in the Hyprland config. Sway and other Wayland setups without the portal need the same kind of binding. The Linux guide shows hold-to-talk on F9 like this:

bindsym F9 exec hex dictation start
bindsym --release F9 exec hex dictation stop

GNOME 46 and older, including Ubuntu 24.04, have no shortcuts portal. Settings can write custom keyboard shortcuts that run hex dictation toggle, hex paste-last, and, if Annotate is on, hex annotate toggle. Those custom shortcuts fire on the key press only, so dictation toggles.

You can change the shortcut, the microphone, the language, and the model in Settings. On a Mac, the menu bar item Transcription Model switches downloaded models without opening Settings.

Models and settings

The first run downloads a speech model. After that, ordinary dictation does not need a network connection for the transcription itself. Model downloads come from Hugging Face and its CDN. App updates come from anoma.ly.

An English system starts on Parakeet Unified English, the catalog default in 3.1.2. You can switch models later, and each model remembers the language you last selected for it. The README’s example is English with Parakeet and French with Whisper, switched from the menu bar.

These are the built-in models in the 3.1.2 catalog. Coverage is the label Hex stores on each one. Streaming means the model can transcribe while you hold the shortcut, on a GPU.

ModelCoverage in the catalogStreams
Parakeet Unified EnglishEnglishYes
Parakeet v2EnglishNo
Parakeet v325 languagesNo
Whisper large-v3-turbo100 languagesNo
Qwen3-ASR 0.6B30 languagesNo
QwenCleo-ASR 1.7BEgyptian Arabic, with English termsNo
ivrit.ai Whisper large-v3-turboHebrewNo
SenseVoice SmallChinese, Cantonese, English, Japanese, KoreanNo
Cohere Transcribe14 languagesNo
Underdog Bark 3.0 ProEnglishNo
Underdog Bark 3.0 CleanEnglish, without filler wordsNo
Parakeet Ultra25 languagesNo
Nemotron 3.5 ASR27 languagesYes
Apple SpeechSystem localesNo

Parakeet Ultra sits behind its own Labs toggle. Apple Speech is ready only on macOS. On Windows and Linux that runtime is off.

With Parakeet Unified English on a GPU, Hex streams audio into the model while the shortcut is held. macOS uses Metal. Windows x64 and Linux use Vulkan. Release then only has to finish the last one or two seconds. Windows ARM64 and other CPU devices transcribe the whole clip. Nemotron 3.5 ASR is the other streaming model in this catalog. If a stream fails, Hex transcribes the whole recording instead.

You can also point Hex at a GGUF file, or at a speech model already in the Hugging Face cache on that computer. Catalog models are the ones Hex pins and tests. A file you add is marked as not tested by Hex. Architectures it will consider are Parakeet, Whisper, Qwen3-ASR, Cohere, and SenseVoice, with F32, F16, or Q8_0 tensors.

Settings also cover recording behavior, feedback volume, launch at login, and, on a Mac, Dock visibility. Corrections in Modes fix names and spellings before the paste. Modes can add optional rewriting and text transformations, per app or per site.

Labs, at the bottom of the sidebar, holds experimental features. The 3.1.1 notes say the toggles added there stay off until you turn them on. hex labs list shows the switches in the running app.

Custom transformations live in ~/.config/hex/hex.config.ts. Hex runs that file in an embedded engine with no filesystem, network, or process access. hex commands init creates a starter, and hex commands check reports what it registered.

Dictate a prompt to a coding agent

Hex doesn’t know the focused app is a coding agent. It pastes into whatever text field has focus, whether that’s Cursor chat, a terminal TUI, or any other box.

So you click the prompt, hold the shortcut, speak one task, and release. Read the text and fix paths and product names before you send it.

Say the task, the boundary, and how you will check it. Paths like src/pages/reading-list.astro are safer typed from the keyboard.

The voice to text guide has a spoken prompt next to the version that should be on screen before you press Enter.

Don’t dictate a token, a password, or the contents of .env. A local transcript still lands in the prompt, and the agent may send that prompt to a hosted model.

OpenCode

Ordinary dictation doesn’t need OpenCode, and Hex never launches it. Optional rewriting uses an OpenCode service that is already running, and if it’s missing, Hex says OpenCode isn’t running.

Turn the OpenCode step on in Modes when you want a rewrite before the paste. Hex then sends the transcript, the foreground application, and, in Brave, the website hostname to the model provider you configured. Website-aware modes currently require Brave. If the rewrite fails, Hex keeps the corrected local transcript and still runs any text transformations you selected.

On a Mac, the history command can filter retained transcripts. This is the example from the README:

/Applications/Hex.app/Contents/MacOS/hex history --since 3d --app 'OpenCode Beta' --grep problems --json

The app has to be running, and the query doesn’t start the microphone.

Claude Code, Codex and Cursor

Claude Code, the ChatGPT desktop app and Cursor each have their own dictation, and the voice to text guide covers how each one works. Hex doesn’t replace them. It pastes into whichever prompt has focus, with the same Option shortcut on a Mac, so it doesn’t clash with Claude Code’s Space, Cursor’s Command-Shift-Space or its Control-M in the Agents Window.

Use Hex when you want one shortcut across the editor, the terminal and the chat. Use the built-in microphone when you’re already in that app and want its transcription. In a terminal, the terminal has to be the focused app and the cursor has to be in the prompt.

How I would use it

If I were dictating prompts with Hex, I would leave OpenCode rewriting off and keep Parakeet Unified English for English. I would speak one task, then type any path the transcript missed. I would not point it at a chat that auto-sends on Enter until I had read the line.

Privacy

According to the README, speech transcription runs locally. Ordinary dictation does not require OpenCode or a cloud transcription service.

Hex does not save dictation audio by default. History keeps successfully pasted text for seven days by default, including the local transcript, the final text, and limited app and timing metadata. It does not store audio, full browser URLs, or window titles. Retention is capped at 10,000 entries and 16 MiB of raw plus final text. Oldest entries go first.

History > Keep: Off stops new entries. Clear all removes the ones you have. Turning retention off does not delete existing entries, and it does not turn off diagnostic logs.

An environment variable, HEX_RETAIN_DICTATION_AUDIO, can keep a bounded number of WAV files for diagnostics. Local logs can contain transcript text, recognized command speech, foreground app names, and Brave URLs. Hex does not upload those logs.

OpenCode features send the transcript and the context described above to your configured model provider. Model installs and app updates need the network.

Where a Hex transcript can goThe speech model runs locally and the paste stays on the machine. OpenCode rewriting is optional and sends the transcript to the configured model provider.OffOnMicrophoneLocal speech modelPaste into thefocused appOpenCode rewrite on?Model provider

Tips

Speak one task. A wandering sentence becomes a wandering prompt.

Pause between sentences if you want them as separate lines. Saying “new line” isn’t documented to insert a break in the default model, and a 3.1.0 note moved spoken formatting, including “new line” and “new paragraph”, into Labs, off until you turn it on. Check Settings › Labs on your build before you rely on that phrase.

Add names Hex keeps missing. Modes corrections are the README’s path. The CLI also has hex vocabulary to list, add, and remove terms, against the running app.

On Linux, pick a dedicated microphone if you use Bluetooth earbuds. Recording from the headset switches it into a hands-free profile and the audio quality drops. The Linux guide recommends the laptop microphone or a USB microphone for input, and the earbuds for output only.

Leave Annotate alone until you want it. It is on by default on macOS and Linux, and not on Windows yet. Its key is Option-Shift-S on a Mac and F9 on Linux. While it records, the dictation shortcut does nothing.

If you manage updates yourself, the README documents DisableUpdates for the anomaly.hex preference domain on macOS, and a DWORD of the same name under HKLM\SOFTWARE\Policies\Anomaly\Hex or HKCU on Windows.

Troubleshooting

The README has a troubleshooting table:

SymptomCheck
Hex appears to disappearOpen Hex from its menu-bar hexagon. If the status item could not start, Hex keeps its Dock icon visible.
The shortcut does nothingOpen Hex and verify Microphone and Accessibility access.
Recording works but text does not pasteVerify Accessibility access.
Transcription does not startConfirm that the selected local model is installed.
OpenCode features failVerify the separate OpenCode installation and model provider.
Website context does not matchWebsite-aware modes currently require Brave Browser.

On Linux, a shortcut that does nothing is often the desktop, not the microphone. GNOME may have refused Alt+Space and be waiting on Ctrl+Alt+Space. Hyprland needs a bind in your config. KDE can keep an older binding after you record a new one. Settings shows what the portal granted.

If the paste never arrives, look at wl-clipboard and wtype, or the RemoteDesktop permission on KDE and GNOME. The transcript should still be on the clipboard.

hex devices lists ALSA inputs. On a machine with no real microphone it can print a dummy device instead of a headset. That output is the device name, not a Hex error string.

Logs are in these folders:

  • macOS: ~/Library/Application Support/hex/logs/
  • Windows: %APPDATA%\hex\logs\
  • Linux: ~/.local/share/hex/logs/, or under XDG_DATA_HOME

Discord on Windows can show “Playing HEX: Shards of Fate” while Hex is running, because it matches hex.exe to that game. The Windows guide says to hide the entry under Discord’s Settings › Registered Games, or turn activity sharing off. Hex keeps working either way.

Report a bug with the version, the operating system, and the steps. The issue template is on the GitHub repository. Redact transcripts and logs before you attach them.

Hex and the other dictation apps

Hex runs on Apple silicon Macs with macOS 15 or newer, Windows 10 and 11, and Linux. Audio stays on the machine unless you turn on the OpenCode rewrite, which sends the transcript to your model provider. It’s MIT licensed, and the download page lists no price.

For Superwhisper, MacWhisper, Wispr Flow, macOS Dictation and the rest, with prices and where the audio goes, see the AI dictation app comparison. Handy, which that guide uses as the daily local app, is the closest cousin to Hex: also free, local, and cross-platform.

FAQ

What is Hex?

Hex is a desktop dictation app from Anomaly. You hold a shortcut, speak, and it pastes a transcript into the focused app. This page covers the 3.1.2 stable release.

Is Hex free?

The download page does not list a price. The GitHub repository is MIT licensed, which allows use without a license fee. There is no subscription in the install docs.

Does Hex work on Linux and Windows?

Yes, as betas, beside the Apple silicon Mac app. Windows is 10 and 11, x64 and ARM64. Linux is x86_64 and ARM64, with glibc 2.39 or later. Intel Macs are not in the download table.

Does Hex send my voice to the cloud?

Speech transcription runs on the machine. The first model download and later app updates use the network. If you turn on OpenCode rewriting, that feature sends the transcript to the model provider you configured. The audio of ordinary dictation is not saved by default.

What is the Hex keyboard shortcut?

Hold Option on a Mac. Hold Ctrl+Space on Windows. Hold Alt+Space on Linux, or Ctrl+Alt+Space on a new GNOME setup. You can change the chord in Settings.

Can I dictate into Claude Code, Codex, or Cursor?

Yes, whenever their prompt field is focused. Hex pastes text. It does not replace Claude Code’s /voice, Codex Dictate, or Cursor’s microphone button. Those are separate controls, covered in the voice to text guide.

Want me to talk about your product? You can sponsor this site.

Related posts about ai: