Dictation Without the Cloud Tax: HEX's Local-First Bet

For two decades, voice-to-text was the technology perpetually five years out. By most accounts, 2026 is the year it quietly arrived. Industry research puts average speaking speed around 150 words per minute against 40–60 for typing, and even after you subtract the time spent fixing what the recognizer mangled, voice still wins by roughly 2.5× — a real margin, not a demo-floor fantasy. Roughly half of U.S. workers now use AI on the job, and a fast-growing slice of that usage is voice rather than keyboard. Nearly two million U.S. workers a year are affected by repetitive strain injuries, which makes dictation less a productivity hack than a way to keep working at all. Into this moment steps HEX, an open-source, local-first dictation app written in Rust, and the question worth asking is not whether voice input is useful — the data has settled that — but whether it needs to live in someone else’s datacenter to work well.

Hold, speak, release

HEX’s core interaction is almost aggressively simple: hold a shortcut, speak, release. The app transcribes on your machine and pastes the result into whatever currently has focus. Double-tap to keep recording, Escape to cancel, a secondary shortcut to re-paste the last dictation. That’s the whole product surface for most users, and the restraint is the point. The category’s winners — the ones Zapier’s 2026 roundup catalogs, from Apple’s built-in Dictation to the cloud-native Wispr Flow — all converge on this same push-to-talk grammar, because it maps cleanly onto how people already think about walkie-talkies and voice notes. HEX’s differentiator isn’t the gesture. It’s where the compute happens.

Transcription runs entirely on-device. During setup the app downloads a local speech model — Whisper or NVIDIA’s Parakeet are the named options — and ordinary dictation never touches a cloud service afterward. On Apple silicon Macs running macOS 15 or newer, this is the shipping product. On Windows and Linux, HEX ships as explicit betas, and the Windows one carries a genuinely impressive technical detail: on x64 machines with a Vulkan graphics driver, Parakeet Unified English streams while you speak, so text lands within about 100 milliseconds of releasing the key even after minutes of continuous talking. ARM64 Windows falls back to CPU transcription. That streaming figure matters because latency is where local transcription has historically lost to cloud services with armies of GPUs — and where a Vulkan-accelerated local model closes the gap without a network round-trip.

The Linux beta is narrower: x86_64 only, targeting X11 and compatible wlroots Wayland compositors, with native Wayland support contingent on the compositor implementing clipboard, virtual-keyboard, and layer-shell protocols. It’s the honest kind of beta — the README tells you exactly which desktop environments it will disappoint.

The boring part is the product

Most dictation apps live or die not on model quality but on the unglamorous plumbing: global hotkeys, clipboard integration, focus tracking, permission wrangling, model management. This is where HEX’s Rust rewrite — it superseded an earlier Swift app, whose settings and history deliberately do not carry over — appears to have spent its effort. The app handles the macOS permission triad (microphone, input monitoring, accessibility) with a guided setup, supports per-model language memory so you can flip between English-on-Parakeet and French-on-Whisper from the menu bar, and ships signed builds with an in-app updater that organizations can disable via a managed preference key when they’d rather control versions themselves. That last detail — a deployment story for fleets of managed Macs — is the kind of thing nobody builds for a hobby project.

The failure handling is similarly considered. If optional post-processing fails, HEX keeps the corrected local transcript rather than pasting nothing. If a transformation stage fails, it keeps the pre-stage text rather than a half-mangled result. A failed voice action pastes nothing at all, which is the correct choice: silence beats garbage in a text field. None of this is glamorous, but it’s the difference between a dictation tool you trust and one you abandon after a week.

Privacy as architecture, not marketing

The privacy section of HEX’s README is unusually specific, which is itself a signal. Speech recognition runs locally. Audio isn’t saved by default. History retains pasted text for seven days with explicit entry and byte limits, never stores audio, full URLs, or window titles, and can be turned off or cleared entirely. Diagnostic logs stay on the device. There’s even a named opt-in setting for retaining a bounded set of owner-only WAV files, for people debugging recognition problems who understand the tradeoff. This granularity — byte limits, retention defaults, what exactly a log file can contain — reads like documentation written by someone who has actually fielded a “where does my voice go” question from a paranoid user, which is the only kind of privacy documentation worth reading.

But HEX is not a purity project. It has an escape hatch, and the README is candid about what leaks through it.

The OpenCode coupling

HEX’s optional features integrate with the OpenCode V2 beta, and the integration is opt-in with an unusual degree of separation: both voice actions and provider-backed rewriting default off, use different shortcuts than dictation, and are described as entirely separate from local speech recognition. When enabled, Voice Action sends your instruction, selected text where accessibility makes it available, the foreground application, and — curiously — the hostname of the current Brave website to your configured model provider. Post-processing sends the transcript plus the same context. The Brave-specific website awareness is a genuine rough edge: application- and site-aware modes currently require Brave Browser, full stop, which is an odd dependency for an otherwise platform-agnostic tool.

The more interesting wrinkle is the custom commands system. Advanced users can write personal commands as TypeScript configuration files, which requires the Bun runtime and — as the README states plainly — constitutes executable local code running with your user permissions. HEX even ships an installable “skill” for coding agents, so you can have Claude or a similar tool write your voice-command configuration for you. There’s something both very 2026 and slightly vertiginous about that loop: an AI agent configuring the voice commands that will later invoke AI actions on your machine. The README’s advice to review agent changes and install only trusted dependencies is correct, and also the entire security model.

Where it sits in the field

The dictation landscape in 2026 is stratifying along a privacy axis. At one end sits Wispr Flow, which by industry reporting has reached 270 Fortune 500 companies and a 70% twelve-month retention — stickiness the Dragon era never achieved — by betting on cloud models and team features. At the other end, Apple Dictation offers free on-device processing but locked to Apple’s ecosystem, and tools like Handy offer open-source, offline, press-and-talk simplicity under an explicit “accessibility tooling belongs in everyone’s hands, not behind a paywall” ethos. Microsoft’s store carries Murmur, another offline speech-to-text entry. HEX occupies a distinctive middle position: local-first like Handy, but with a cross-platform ambition, a commercial-grade update and deployment story, and an optional AI layer that Handy deliberately refuses to build. It’s closer to what Wispr Flow would be if Wispr Flow respected your bandwidth bill.

The legacy comparison is instructive, too. Dragon Medical One has anchored clinical dictation for over a decade, and healthcare systems’ retrospective analyses of it read as a reminder that domain-specific dictation is a serious, high-stakes market. HEX is not competing there — yet. But the underlying economics are shifting: as accuracy benchmarks for general models climb into the mid-90s percentile across accents and languages, the moat around specialized vocabularies gets shallower, and a local model that never sends patient audio anywhere starts looking like a compliance feature rather than a limitation.

Rough edges, honestly catalogued

The README’s candor extends to its own deficiencies, which is refreshing. The Windows installer isn’t code-signed yet, so SmartScreen will warn — the documented workaround is the standard “more info, run anyway” dance that every unsigned Windows binary forces on its users. The TypeScript SDK is published on npm, but the bundled native helper isn’t, meaning developers must supply a compatible HEX executable themselves; the SDK is more promise than product at this stage. Voice commands are experimental, macOS-only, and require a separate local model. The Linux story excludes voice commands and meetings entirely. And the name itself is a discoverability disaster: search for “Hex” and you’ll find an AI analytics platform, a classic board game, a color-code reference site, and a dictionary entry before you find a dictation app. Community discussion exists — a Reddit thread on offline dictation software is among the results — but its contents are walled off, so it’s unclear how much grassroots traction HEX has versus how much is still potential energy.

The bet, and whether it pays

HEX’s wager is that the cloud dictation incumbents have a structural weakness: their best features — context-aware rewriting, voice-driven actions — depend on shipping your words to a provider, and a meaningful population of users will always balk at that, whether for privacy, compliance, or the simple indignity of renting back their own voice. The local-first alternative has historically lost on quality and latency. With Whisper-class models running on consumer silicon and streaming transcription landing within 100 milliseconds, that excuse is evaporating.

What remains unresolved is the AI layer. HEX’s answer — local recognition as the always-on foundation, provider-backed intelligence as an explicit, separately-toggled opt-in with documented data flows — is a reasonable architecture, but it means the most hyped capabilities are precisely the ones that break the local-first promise. The project’s future likely depends on whether local models grow good enough at rewriting and instruction-following to close that gap, or whether HEX remains forever a superb transcription engine with a cloud-shaped asterisk. Either way, it’s a serious piece of systems software — MIT-licensed, Rust-built, honest about its limits — in a category that suddenly everyone cares about. That combination is rare enough to watch closely.