Image: GitHub
Why it mattersBuyers of a voice tool often pay for the cloud transcription too. Hex is a free, local option. A team can test whether the vendor cost is still worth it.
A free, local alternative to a paid voice-typing tool just reached a wider audience. Hex is an MIT-licensed desktop app by developer Kit Langton, under the GitHub user anomalyco. The GitHub API reported 724 stars for anomalyco/hex on 29 October. The repository was created on 17 July 2026. Trendshift ranked it 22nd on its daily list the same day.
The pitch, in the README's own words: "Hold a shortcut, speak, and paste into any app." The app runs on-device, so the audio never leaves the machine. The code is Rust. The latest release is v2.3.1, published on 21 October 2026.
What it does
Hex sits in the macOS menu bar. The user presses a configurable shortcut, speaks, and Hex writes the transcript into whichever app has focus. The README names macOS on Apple silicon as the main target. Windows x64, Windows ARM64 and Linux are also listed. Linux is in beta.
The speech model runs locally. The README offers several model sizes. A user can pick between speed on a small Mac and accuracy on a larger one. There is a history pane with privacy controls. An optional "Voice Action" mode can rewrite or clean up the transcript before pasting it. A TypeScript SDK lets other apps call Hex.
What a reader can do with it
A team with people who dictate for notes, email or chat has a free, local option. They can compare it to the paid cloud tools. The install needs no key and no subscription. For a user who writes plain prose but cannot send audio to a cloud, local transcription is the whole product.
For a user who already pays for a voice tool, the question is simple. Is on-device accuracy on small models close enough now that the cloud step is not worth the cost? Hex is the free way to find out.
Hex also matters as an agent-adjacent tool. A coding agent that drives a Mac can call the SDK. A short voice note turns into pasted text. The user does not have to type out the first draft of a prompt. For a local AI assistant, Hex keeps one more step off the network.
What the project does not do
The README does not list a benchmark against the paid tools. There is no accuracy figure for any one model, no word-error-rate on a shared dataset. A buyer should run their own voice through Hex for a day and judge from that.
The TypeScript SDK is new. There is no published list yet of apps calling it. And the Linux beta is beta. A Linux-only team should expect rough edges.
Source
- anomalyco/hex, GitHub
- Hex 2.3.1 release, GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
