Privacy.
Whatever you select, whatever you summarise, whatever your Claude sessions produce — none of it touches a server. The text starts on your Mac and ends on your Mac.
Myna lives in your menu bar and reads any selection, any article, or any finished Claude Code session aloud with a single hotkey. Everything happens on your Mac. Nothing ever leaves it.
Some afternoons the screen turns to gauze. The words you've read since morning blur into one long ribbon, and the prose you still owe the day feels heavier than it should. Myna is for those afternoons — a small companion in the menu bar that takes the reading off your shoulders and gives it back to you as a voice.
Highlight any text anywhere on your Mac and press ⌘⌥⇧S. A warm, MLX-rendered voice picks it up midstream and reads it back to you. Need the gist instead of the whole essay? ⌘⌥⇧A hands the selection to a local Qwen model and speaks a summary — in your own room, on your own silicon.
On Listening
There is a particular hour of the afternoon when the screen begins to gauze over, when the prose you owe the day grows heavier than it should, and you find yourself reading the same paragraph for the third time. That is the hour Myna was built for.
Open an article in Chrome, hit ⌘⌥⇧R, and Myna pulls the main body out of the page and starts reading. No sidebar clutter, no cookie banners, no advertising voiceover. Just the piece, the way it was meant to land.
If you run parallel Claude Code sessions, you know the chaos of them all finishing at once. Myna quiets that. As each session completes, it whispers itself into the menu bar and waits. You click the one you want to hear. The others stand by, patient, until you're ready. No talking over each other. No missed answers buried in noise.
Real audio, not afplay: AVAudioEngine drives playback, so speed changes don't pitch-shift and you can scrub or jump ±15s mid-sentence. A native Settings panel rebinds every shortcut, picks the voice, and points the daemon. And myna:// URLs let BetterTouchTool, Shortcuts, or Alfred drive Myna without simulating a keystroke.
The voice model (Kokoro, af_heart) runs on your machine. The summariser (Qwen 3.5 4B via Ollama) runs on your machine. No API key, no usage meter, no telemetry. MIT-licensed and open at github.com/PrerakGada/myna. What you read stays with you, and it stays free.
Each part does one job. The engine speaks. The brain decides what to speak. The surface lets you press a key. None of them call the internet.
Native SwiftUI · AVAudioEngine
The menu bar, the hotkeys, the Settings panel, the playback. AVAudioEngine handles speed without pitch shift and lets you scrub or jump ±15s. Signed, notarised, and quietly updated by Sparkle.
Python · FastAPI · streams WAV
A small local service that synthesises in chunks and streams them to the app as the voice plays. Extracts articles, summarises through Ollama, never opens a socket beyond 127.0.0.1.
mlx-audio · Kokoro af_heart
The Kokoro model running natively on Apple Silicon. Fast because it's local. Warm because Kokoro is just a genuinely good model.
Whatever you select, whatever you summarise, whatever your Claude sessions produce — none of it touches a server. The text starts on your Mac and ends on your Mac.
Cloud voices charge per character, per minute, per month. Myna charges nothing, and will charge nothing, because there's no one to charge you.
A round trip to a TTS API is a beat you can feel. Local inference on Apple Silicon doesn't have that beat. You press the key, the voice starts.
Myna ships as a native, code-signed, notarised macOS app. Install with Homebrew or grab the DMG from GitHub Releases — either way, the local voice daemon comes along for the ride. macOS Ventura or later, Apple Silicon only.
Defaults use ⌘⌥⇧ so they don't collide with the shortcuts you already love. Rebind from the menu bar at any time.