Talk to your coding agents. Keep every word on your Mac.
Realtime, fully local dictation for the menu bar. Press a key and speak. Your words appear while you're still talking.

Install · Documentation · Coding agents · Contributing

localvoxtral streams text as the audio arrives instead of transcribing after you stop speaking. It runs open streaming speech models on your own Apple Silicon: Mistral AI's Voxtral Mini 4B Realtime by default, or NVIDIA's Nemotron 3.5 ASR Streaming on Macs with less memory.

It is built first for prompting coding agents by voice, and it works as a general dictation app in any other app too. Everything runs on-device, with no account and no subscription. Nothing leaves your Mac unless you point it at a server yourself.

curl -fsSL https://raw.githubusercontent.com/T0mSIlver/localvoxtral/main/scripts/install.sh | bash

Or install with Homebrew:

brew install --cask T0mSIlver/localvoxtral/localvoxtral

You can also download the latest DMG from Releases. localvoxtral needs an Apple Silicon Mac on macOS 15 or later. The local engine downloads about 6 GB on first setup and runs best with 16 GB of memory; on an 8 GB Mac, turn polishing off or pick the smaller models in Settings → Engines.

On first launch, a setup wizard asks for permissions and downloads the engine.

• Streams as you speak. Your words appear while you're still talking, transcribed on your Mac. No telemetry, no account, no cloud fallback (how it works).
• One key. Tap or hold Fn/Globe to dictate into an overlay you review before it commits, or type live into the focused app (shortcuts).
• Keeps your words. Optional LLM polishing fixes punctuation and misheard words, and leaves the rest as you said it. Fix a misheard name once before you send a prompt, and the app spells it right in that project from then on (your terms).
• Built for coding agents. Dictate into Claude Code, Codex, opencode and Mistral Vibe, in any terminal, a herdr or cmux pane, over SSH, or a claude.ai/code tab. Dictation joins the Claude Code session under your cursor, and polishing reads its screen and the files it touched. "dash dash force" becomes --force (details).
• Know when an agent needs you. When a coding agent waits for an answer, localvoxtral shows a banner, plays a sound and marks the menu bar icon (turn on Tell me when an agent needs you in Settings → Dictation) (details).
• Hosted models, or your own server. With one Mistral key, dictation can run on Voxtral Realtime 3 and polishing on Mistral Medium 3.5 instead, and either can point at any OpenAI-compatible endpoint. Only then does anything leave your Mac (details).
• Multilingual. Dictate in English, French, or any other language your speech model understands. Polishing answers in the language you spoke (English and French are covered by the test suite).

Every guide is listed in the documentation index.

MIT