Hex is push-to-talk dictation that runs on your own machine. Hold Option, say what you mean, let go, and Hex pastes the text into whatever app has focus. The speech model runs locally, so ordinary dictation needs no account, no cloud service and no subscription.
It's free and MIT-licensed, written in Rust, and made by Anomaly (anoma.ly), the team behind OpenCode. Version 3.0.0 shipped on 31st October 2026.
The promo clip from the Hex site, by Anomaly.
What it does
| Key | Action |
|---|---|
| Hold Option | Record while held, transcribe and paste on release |
| Double-tap Option | Hands-free: keep recording until you press Option again |
| Esc | Cancel the recording, or the dictation still transcribing |
| Option-Shift-V | Paste the last dictation again |
| Option-Shift-S | Annotate: record the screen while you draw and talk |
The shortcut is configurable, including left or right modifiers and both Option keys together. On Windows the default is Ctrl+Space.
Beyond plain dictation:
- Vocabulary. List names and terms, and Hex steers the transcription toward them and fixes common mishearings.
- Modes. Corrections and replacements per app, applied before the paste. With Brave, modes can also target websites.
- Transformations. A
~/.config/hex/hex.config.tsfile holds your own text transformations in TypeScript or JavaScript. Hex runs it in a sandboxed engine with no file, network or process access, and reloads it on save. - OpenCode rewriting (optional). If an OpenCode service is running, a mode can send the transcript to your model provider to rewrite it. If that fails, Hex pastes the local text.
- Annotate. Record the screen, draw on it and talk, then hand the keyframes, drawing and transcript to an agent. On macOS and Linux.
- Meetings (beta, macOS). Records and transcribes meetings, in the Labs section.
- History. Pasted text, kept 7 days by default, never the audio.
How it works
- Models: you download a speech model during setup. NVIDIA Parakeet and OpenAI Whisper are both available, and the menu bar switches between them, each with its own language. For example: English with Parakeet, French with Whisper.
- Streaming: with a Vulkan graphics driver on Windows and Linux, Parakeet transcribes while you speak, so text arrives about 100 ms after you release, even after minutes of talking.
- Paste: through Accessibility on macOS. No Input Monitoring permission.
- CLI: the app binary answers read-only queries while Hex runs:
/Applications/Hex.app/Contents/MacOS/hex history --since 3d --json
/Applications/Hex.app/Contents/MacOS/hex meetings show latest --md
- Agent skill: Anomaly publishes a skill so a coding agent can write your transformations:
npx skills add anomalyco/hex --skill hex-personal-commands.
Platforms and install
| Platform | Status | Requirement |
|---|---|---|
| macOS | Stable | Apple silicon, macOS 15 or later |
| Windows | Beta | Windows 10 or 11, x64 or ARM64, not yet code-signed |
| Linux | Beta | x86_64 X11 or wlroots Wayland; Debian, Fedora, Nix, AUR |
On a Mac:
brew install --cask anomalyco/tap/hex
Then grant Microphone and Accessibility, and let setup download a model. Model downloads come from Hugging Face.
Privacy
- Transcription runs locally. Audio isn't saved by default.
- History stores text and limited app and timing metadata, never audio, full URLs or window titles. Retention can be turned off.
- Local logs can contain transcript text. Hex doesn't upload them.
- Only the optional OpenCode features send text to a model provider.
History
Kit Langton wrote the original Hex as a Swift app for macOS. Its repo, with about 2,900 stars, is now archived and points to the new one. The new Hex is a Rust rewrite under Anomaly, which added Windows and Linux. The anomalyco/hex repo started in July 2026 and has about 1,500 stars.
3.0.0 starts fresh. It's a new app ID (anomaly.hex), signed by Anomaly. It imports nothing from 2.3.1: re-grant the macOS permissions and download the model again. 3.0 also dropped spoken commands and Voice Action.
Hex vs the alternatives
| App | Runs | Price | Source |
|---|---|---|---|
| Hex | Local | Free | Open, MIT |
| Superwhisper | Local or cloud | Free, Pro $8.49/mo | Closed |
| OpenWhispr | Local or cloud | Free up to 2,000 words/week, Pro $8/mo | Open |
| macOS Dictation | Local | Built in | Closed |
| Dictee, my own app | Local | - | Private |
My own notes: see Superwhisper — AI Voice-to-Text for Mac, Windows & iOS, OpenWhispr – Open Source AI Voice Dictation and macOS Voice Dictation Solutions for the field, and Dictee: my own local dictation app for the app I built.
My take
Hex is the cleanest free option for local push-to-talk dictation, and the only one that covers Mac, Windows and Linux from 1 codebase. The open source and the sandboxed transformations make it easy to trust.
For me it's a reference rather than a switch. Dictee already does the same job on my Mac Studio, with Parakeet on MLX and a local cleanup model. 3 Hex ideas are worth borrowing for Dictee: the vocabulary list, Annotate for showing an agent what I mean on screen, and a history CLI an agent can query.
The main limitation: rewriting needs OpenCode running. Without it, Hex pastes the raw transcript with your corrections, not a cleaned-up version.
Further reading
The source code:
The release notes for 3.0.0:
The optional companion for rewriting:
