What happened
Anomaly, the team behind OpenCode, published Hex 3.0.0 on GitHub on October 31. Hex is a voice-dictation app: a user holds a shortcut, speaks, and the recognized text is pasted into the focused application. The release describes speech recognition as local. The release announcement presents Hex as a tool for talking to coding agents, though the dictation workflow can send text into whatever app is focused. Downloads are directed from the Hex site; the release material lists macOS, Windows and Linux support, including Linux ARM64 builds and .deb and .rpm packages. No price is stated in the supplied release notes.
Version 3.0 consolidates settings across macOS, Windows and Linux. The sidebar groups General, Output and History. Output includes replacements, per-app modes, OpenCode rewriting and transformations. A recording indicator is available on every OS. Users can build a vocabulary of names and terms so Hex can steer transcription toward them and correct common mishearings. Shortcut configuration now supports both Option keys together, an individual left or right modifier, or a mouse button. The release also adds selectable feedback sounds. On macOS, Hex can lower other audio during dictation, and pausing media now pauses the browser too.
The release says local models use about 160 MB of memory, compared with about 900 MB previously. It adds the Underdog Bark 3.0 Pro and Clean speech models, as well as Bangla support; the model picker displays each model’s accuracy. The software skips recordings with no speech instead of transcribing them. The Hex command-line interface can control the running app, including status, dictation, settings and meeting recording, and can read history and meetings. Hex now has one updater across operating systems, with Stable and Dev channels in settings. Linux receives ARM64 builds, .deb and .rpm packages, and a tray icon.
Some capabilities are explicitly limited by platform or release channel. Meetings is a beta feature in Labs on macOS: it can remind users before calendar meetings after a calendar is connected, record calls from a selected app, and finish a recording on the next launch if Hex quit during a meeting. Annotate is available in Labs on macOS and Linux Dev builds; it records the screen while the user draws and talks, then sends that material to an OpenCode session. Meetings and Annotate should not be read as generally available features across all platforms.
The upgrade is deliberately non-migrating. Hex 3.0 stores data in a folder named hex rather than voice-control and imports nothing. Settings, history, meeting recordings and downloaded models therefore start empty, and the first launch downloads a speech model again. Previously exported meeting notes in ~/Documents/Hex Meetings remain in place; the old data folder is left untouched and can be deleted. Version 2.3.1 does not update to 3.0 automatically. On macOS the app has a new identifier, anomaly.hex, and users must grant permissions again; on Windows, the old app remains installed until removed manually. The release also removes Voice Action and its spoken commands.
What the archive adds
A prior Telegram post described OpenCode working with Docker Model Runner through an OpenAI-compatible interface to run local models without cloud APIs. Hex is a new release in that broader local-builder-tool thread, not a new fact about that integration: it adds local speech input rather than a model runtime. No earlier Field Note supplied here directly covers Hex.
Trend signal
Hex offers a small example of the harness architecture outliving model choice: local dictation, vocabulary, shortcuts and output transformations sit around the coding agent rather than being tied to one named model. The release does not establish adoption or prove that this architecture is winning; it shows one practical interface layer being shipped alongside OpenCode.
Why now
The release itself explains a concrete transition: Hex 3.0 starts a fresh data store, while 2.3.1 does not upgrade automatically. That makes this release a point at which users must choose whether to reinstall and rebuild their configuration, rather than simply receiving an update in the background.
Why it matters
Voice input can reduce the friction of composing long instructions, names or contextual explanations for a coding agent, especially when hands-free input is useful. Local recognition also gives teams a way to keep the audio-processing step on the device, although the release notes do not describe every privacy property or guarantee that all downstream applications handle pasted text locally.
The more operationally significant parts may be vocabulary and per-app output modes. Dictation quality is not just a model problem: project names, technical terms and application-specific formatting affect whether spoken input is usable. A vocabulary that corrects common mishearings, combined with transformations and OpenCode rewriting, gives users controls at the boundary between speech recognition and the agent interface. The claim of roughly 160 MB model memory use may make the tool easier to run alongside development software on constrained machines, but the notes provide no measurement method or device profile.
For teams evaluating it, the cost is not only installation. The clean start means settings and downloaded models must be recreated, macOS permissions must be granted again, and Windows users need to remove the previous app separately. A pilot should account for that reset and distinguish generally available dictation from beta Meetings and Dev-only Annotate.
Limits and uncertainty
The release notes do not provide a price, supported language list beyond the mention of Bangla and Ukrainian interface language, comparative transcription results, or a benchmark method for the memory figures and model accuracy display. They do not quantify latency, battery impact, or how recognition performs in noisy environments. “Local” describes speech recognition in the release, but does not by itself establish the complete data flow for every feature, particularly calendar-connected Meetings or an Annotate session sent to OpenCode. The notes also do not establish that the beta and Dev-channel features are suitable for production use.
Reader takeaway
Hex 3.0 makes local dictation a more complete cross-platform companion for coding workflows, with practical controls for terminology and output. Treat it as a fresh installation: settings, history, recordings and downloaded models do not migrate, and the older version will not upgrade itself. Check platform and release-channel limits before depending on Meetings or Annotate.