Appearance

Language

Mastering AI agents

Hex: local voice dictation from the OpenCode team. Hold Option, speak, and it pastes

Hex is an open-source dictation app from Anomaly, the team behind OpenCode and SST. Hold the shortcut, speak, and the text is pasted wherever your cursor is. Speech recognition runs on your own machine; it is MIT licensed and runs on macOS, Windows and Linux. For Chinese it recommends Qwen3-ASR 0.6B; the Mac app needs Apple silicon and macOS 15 or later.

Kelvin Wang · · 6 min read

On a Mac the shortcut is Option: hold it, speak, let go, and an editor, a browser, Slack or a terminal gets the text. The first run downloads a speech model, and after that your audio does not leave the machine. The source is public on GitHub, with installers for each platform.

GitHub stars (2026-11-05)
5,127
github.com
Latest stable release (2026-11-02)
3.3.0
github.com
License
MIT
github.com

From a Swift app to a Rust rewrite

Hex started as Kit Langton’s personal macOS app, written in Swift, at kitlangton/Hex (created January 2025). That repo is now archived and points to anomalyco/hex.

The new version started in July 2026 as a Rust rewrite, one codebase for three platforms. Old users don’t have to migrate by hand: both versions sign updates with the same Sparkle key, so the old app is offered the new one as an ordinary update. Settings, history and downloaded models do not carry over, and macOS permissions have to be granted again. The new app doesn’t support Intel Macs or macOS 14; those users can stay on the Swift version.

Using it

Install:

  • macOS (Homebrew): brew install --cask anomalyco/tap/hex
  • Linux: curl -fsSL https://anoma.ly/hex/install-linux.sh | sh, or .deb, .rpm, AUR (yay -S hex-bin), Nix
  • Windows: the x64 or ARM64 installer from the website

On first launch it walks you through Microphone and Accessibility permissions (Accessibility is what lets it paste), then downloads a speech model. If the download fails or stalls, you can skip it and pick a model later.

There are three ways to dictate:

  • Hold: hold the shortcut while you talk; release to transcribe and paste. The default is Option on Mac, Ctrl+Space on Windows, and Alt+Space on Linux (Ctrl+Alt+Space on GNOME and KDE).
  • Double-tap to lock: for longer dictation without holding the key; press again when you’re done.
  • Toggle: press once to start and once to stop; Esc throws the recording away.

The shortcut can be any modifier, a key combination, Fn, or a mouse side button (on Mac, turn on Labs › Mouse shortcuts first). If a dictation lands in the wrong place, Paste last dictation puts it where you are now, and every result stays in History.

Speech models, and which one to use for Chinese

Hex’s engine is transcribe.cpp, which runs several open model families. NVIDIA’s Parakeet is the English default; Alibaba’s Qwen3-ASR is the default for Mandarin, Cantonese and Korean; Whisper large-v3 turbo is the default for Japanese and for automatic language detection; GigaAM covers Russian.

For Chinese, the model picker offers three (as set in the source):

  • Qwen3-ASR 0.6B: Recommended
  • SenseVoice Small: Fastest
  • Whisper large-v3 turbo: Reads your vocabulary

The Hex team keeps transcribe.cpp’s benchmark numbers in the repo, scored on the FLEURS test set (Google’s multilingual read-speech corpus) using the same Q8_0 files Hex downloads. Character error rate for Mandarin (CER, lower is better):

ModelMandarin CERLabel in Hex
Qwen3-ASR 0.6B7.57Recommended
Qwen3-ASR 1.7B7.14Not in Hex
Whisper large-v37.98Not in Hex
Whisper large-v3 turbo8.50Reads your vocabulary
SenseVoice Small10.12Fastest

Source: anomalyco/hex docs/research/language-recommendations-other.md (2026-10-22). FLEURS test, Q8_0 GGUF, transcribe.cpp harness.

The Hex team spells out the limits of this comparison. FLEURS is read-aloud Wikipedia sentences, about 900 of them for Mandarin, with a 95% confidence interval of roughly ±0.9 CER points, so gaps under about 1.5 points are noise. The top four models are effectively tied, and Hex offers only two of them; SenseVoice is clearly behind. The harness also doesn’t normalise Traditional and Simplified characters, which can inflate error rates; the doc’s main example is Cantonese.

Names and product terms go in the Vocabulary list. Qwen3-ASR and Whisper both read it as a prompt while transcribing; SenseVoice Small ignores it. Parakeet models use it only with Labs › Vocabulary biasing on, and read at most the first 64 terms. Each term can also carry “sounds like” spellings. Rewriting and search use those, and so does Parakeet’s biasing, but they are not sent to Qwen or Whisper.

Beyond dictation

Modes (Output in the sidebar). The transcript can go through a pipeline before it is pasted: replacements, corrections, transformations, and an optional AI rewrite through OpenCode. Modes switch on the current app or website, and each Mode can have its own vocabulary.

Meetings (beta). Off by default; turn it on in Labs. It is on all three platforms, with calendar reminders on macOS only. It records the other side (system audio) and you (microphone) as separate tracks, transcribes them locally, and saves a Markdown transcript with You and Them turns. It detects calls in Zoom, Google Meet, Teams, Slack huddles and others, and saves to ~/Documents/Hex Meetings by default. On Mac, summaries use Apple Intelligence by default.

Annotate. Record the screen while you draw on it and talk. Hex bundles the video, the ink and a timestamped transcript and sends it to an OpenCode session. It suits pointing at the screen and telling a coding agent what to change.

CLI and MCP. hex status shows the app’s state, hex dictation toggle lets your own keybinding trigger dictation, and hex mcp runs a read-only MCP server so an agent can read past transcripts and meetings.

Labs. Unfinished features sit in Labs at the bottom of the sidebar, each off by default. The Labs doc says features people asked for, which the team built but isn’t sure about, land there instead of shipping on.

Since 3.3.0 the interface is also available in Traditional Chinese. The docs note it was generated from the Simplified strings by a converter and hasn’t been reviewed by a native speaker yet.

No pull requests

The README’s contributing section is two sentences: “We don’t take pull requests. Open an issue describing the bug or idea, and our agents usually turn it around fast.”

The commit history shows the process. In the latest 68 commits, 67 titles carry an issue or task number. AGENTS.md at the repo root is named the implementation authority, and each feature in docs/ has a map of how users reach it, which check proves it works, and what is still unverified. Between November 1 and 2 (UTC) the team shipped five releases, 3.1.0 through 3.3.0.

The README says their agents usually handle reported issues fast. The cost is that you can’t send a fix back upstream; the MIT license lets you fork and change it yourself.

Current limits

  • Mac means Apple silicon on macOS 15 or later. Intel Macs have only the old Swift version.
  • Windows and Linux are betas. Linux supports X11, wlroots-based Wayland compositors, GNOME and KDE; on KDE you approve Hex’s shortcut in the desktop’s settings.
  • “Nothing leaves your machine” covers speech recognition. OpenCode rewriting, OpenCode meeting summaries and Annotate all hand content to whichever model you set up in OpenCode, and if that is a cloud model, the content is sent to it. To stay fully local, leave all three off.

Who it suits

If you want on-device recognition on all three desktop platforms from an open-source app, Hex covers all of that. It fits best if you already use OpenCode, since Modes’ AI rewrite and Annotate plug straight into it.

For Chinese, the default Qwen3-ASR 0.6B is the place to start, with the names you use most in the vocabulary list. If those still come out wrong, try Whisper large-v3 turbo.

Sources

  1. 1. Hex website · Anomaly · checked 2026-11-05
  2. 2. anomalyco/hex (README, docs/, source) · GitHub · commit 2595bc5, 2026-11-05
  3. 3. Hex releases · GitHub
  4. 4. Hex LICENSE (MIT) · GitHub
  5. 5. Hex 3.3.0 release notes · GitHub · 2026-11-02
  6. 6. Language recommendations: zh, yue, ja, ko, the Qwen set, and Auto · GitHub · 2026-10-22
  7. 7. kitlangton/Hex (archived Swift version) · GitHub
  8. 8. anomalyco/opencode · GitHub

About the author

Kelvin Wang

Electrical engineer building tools with AI agents, sharing hands-on notes on engineering and investing.

More about me