Always
Speak. It types. Anywhere.
The always-on dictation app. No push-to-talk, no window to open — just talk.
Website · Download · Install · Features · Shortcuts · Privacy
✨ What it is
You speak. Your words appear at your cursor — in any app, any text field. Nothing to press, nothing to click, nothing to remember. Always sits in your menu bar, listens, and types what you say.
It's less a speech-to-text tool than an idea: your hands should be free, speaking should feel as natural as thinking, and the technology should disappear so completely you forget it's there. That's hard to explain in words. Try it for a day — you'll get used to it faster than you expect.
Other dictation tools want to be the center of attention. Always wants to disappear.
| 1 · Talk | 2 · Always listens | 3 · Your words land |
|---|
| Just start speaking. No hotkey, no button. | A small floating badge shows it's listening and previews your words live. | The text is typed at your cursor, then sent after a 4-second countdown. Press any key to cancel. |
🎯 Features
Everything below is the macOS app. On Linux and Windows, Always runs as a background daemon you drive from the command line — see Install.
📖 Day one, it already knows your words
The first time you open Always, before you've said a single word, it quietly reads the names of the apps on your Mac, your macOS Text Replacements, and the dictionaries you spent months building in Wispr Flow and SuperWhisper — and makes them its own. So the very first time you say "Notion", "Kubernetes" or that client's name you once taught Wispr Flow, it comes out right. Switching costs you nothing you already taught another app.
🗣️ Hands-free
- Always on — no push-to-talk. Start speaking and your words appear.
- Types anywhere — at your cursor in any app: Xcode, Slack, Notion, Terminal, your browser.
- Auto-enter, with a safety net — after typing, Always presses Return following a 4-second countdown. Any key cancels it; ⌃⌥A turns it off.
- Edit with your voice — "scratch that", "change Tuesday to Wednesday", "new line", "make that a list", "make that more formal", "undo".
- Snippets — say a trigger phrase and Always types your saved text, word for word.
🎤 Accurate
- My Voice — record three short samples and Always listens only to you: not the people around you, not the call on your other screen. Your voiceprint never leaves your Mac.
- Talk over music and video — echo cancellation removes your Mac's own sound from the mic. With My Voice on, you never have to pause a video to dictate.
- Learns your vocabulary — correct a word once (⌃⌥W) and it stays fixed. Add your project's jargon and "Kubernetes" stops coming out as "cuber netties".
- Hallucination filter — drops what speech models invent out of silence: "thanks for watching", subtitle credits, repeated words.
🔐 In your control
- You choose where it types — Always starts paused in every app. Turn it on for the apps you want (⌃⌥P), so text never lands in the wrong window.
- On-device by default — on macOS 26+, Apple's speech engine transcribes on your Mac: no account, no API key. Prefer the cloud? Switch to Groq.
- Plays nice with other mic apps — pauses when Zoom, Superwhisper or anything else takes the microphone, and resumes when it's free.
- Nothing gets lost — whatever Always holds back is kept, and ⌃⌥V pastes it.
⚡ Shortcuts
| Shortcut | What it does |
|---|
| ⌃ ⌥ P | Turn Always on or off for the current app |
| ⌃ ⌥ ⇧ P | Master pause — everywhere at once |
| ⌃ ⌥ A | Auto-enter on / off |
| ⌃ ⌥ V | Paste the last transcript Always held back |
| ⌃ ⌥ W | Correct the last transcript — Always learns the fix |
| ⌃ ⌥ X | Learn from your edit — select the text you fixed, press it, and Always compares it with what it typed |
| Hold ⌘ as the text arrives | Skip this paste (it's kept — ⌃⌥V pastes it) |
| Hold ⌥ while speaking | Keep the sentence open through a long pause |
Change them in Settings → Shortcuts on macOS. Linux supports the configured global shortcuts in X11 sessions; Wayland and Windows users can use always toggle-pause and always toggle-auto-enter from the CLI.
🚀 Install
macOS (14+, Apple Silicon) and 🐧 Linux (x86_64 or arm64, glibc 2.39+):
curl -fsSL https://raw.githubusercontent.com/LivioGama/always/main/scripts/install-always.sh | bash
🪟 Windows (10 or 11, x64):
irm https://raw.githubusercontent.com/LivioGama/always/main/scripts/install.ps1 | iex
On macOS you get the full app, and its first launch walks you through everything. On Linux and Windows you get the always daemon and CLI: add a free Groq key and start it (details below).
🤖 Let your agent install it
Paste this into Claude Code, Codex or any coding agent with a shell:
Install Always (https://github.com/LivioGama/always) for me: follow the
"For AI agents" section of its README, then tell me what you did and
what is left for me to click.
macOS, first launch
- Grant Microphone, Accessibility (to type into other apps) and, on macOS 26+, Speech Recognition. Each has a button that opens the right prompt.
- On macOS 26+, Apple's on-device speech model downloads in the background. No account, no API key.
- Optional: record three short samples for My Voice.
- Pick the apps Always types into. Every other app stays paused until you turn it on with ⌃⌥P.
On macOS 14–25, or to use the cloud: add a Groq key in Settings → Models and choose Use Groq (Remote).
🐧 Linux
always config set groq_api_key <your-groq-key>
systemctl --user enable --now always-daemon.service
Or transcribe on your own CPU, with no key and nothing leaving the machine:
systemctl --user enable --now always-daemon.service
always model list # the catalogue
always model download moonshine-medium-streaming-en # English; nemotron-3.5-asr-streaming-0.6b is multilingual
always model use moonshine-medium-streaming-en
Wayland setup (Hyprland/Omarchy, Sway, GNOME, KDE): see docs/SETUP.md.
🪟 Windows: allow microphone access (Settings → Privacy & security → Microphone → Let desktop apps access your microphone), then in a new terminal:
always config set groq_api_key <your-groq-key>
always run
Manual downloads: every installer and package (.tar.gz, .deb, .rpm, .zip), each with its .sha256 and a minisign .minisig signature, is on Releases. The public key is resources/update-signing.pub.
Updating: Always tells you when a newer version exists and never installs it by itself. Run always update (always update --check only reports: exit code 10 means an update is available). It downloads the release, verifies every file against the signing key built into the app, then installs it. On Linux, a .deb or .rpm install is updated this way from a terminal (it asks for sudo); the macOS app also offers an Install update button.
Build from source (macOS): needs Rust and Xcode 26 (older SDKs build without on-device Apple speech). No Homebrew, no SoX: everything Always needs ships inside the app.
git clone https://github.com/LivioGama/always.git && cd always
cargo build --release --bin always
ALWAYS_BUILD_PROFILE=release ALWAYS_SWIFT_CONFIGURATION=release ./Always/build.sh # installs /Applications/Always.app
🔧 Settings
On macOS, everything lives in one Settings window, opened from the menu bar:
| Panel | What's inside |
|---|
| General | Sound cues, status overlay, the apps Always types into |
| Models | Speech engine (Apple on-device or Groq), Groq API key, language |
| My Voice | Record or re-record your voiceprint |
| Permissions | Microphone and Accessibility |
| Behavior | Silence timing, auto-enter, live preview, idle pause |
| Shortcuts | Every shortcut, editable inline |
| Library | Vocabulary and snippets |
| History | Recent transcripts (when transcript logging is on — see below) |
| About | Version and updates |
🔒 Privacy
- No telemetry, no analytics, no account.
- On-device by default — on macOS 26+, Apple's speech engine transcribes on your Mac.
- Groq only if you add a key — audio goes to Groq when it's your speech engine, and transcript text when Groq does the clean-up pass. Without a key nothing is sent; the daemon only opens an empty warm-up connection to
api.groq.com at startup.
- Other traffic — on every OS, an update check: a plain request to GitHub Releases when the daemon starts and then once a day, which sends nothing about you (just a
User-Agent with the app version) and only tells you a newer version exists. Turn it off with always config set update_check false or ALWAYS_UPDATE_CHECK=0. Plus one-time model downloads.
- Your voiceprint stays local — in
~/Library/Application Support/always/voiceprint.json, never uploaded.
- Transcripts stay out of the logs in release builds, unless you opt in with
ALWAYS_LOG_TRANSCRIPTS=1.
- Your API key is stored in Always's local database (not the Keychain) or read from
GROQ_API_KEY, and is never written to logs.
🤖 For AI agents
If you are an AI agent setting Always up for a user, do this, in order. Only install it if the user asked you to.
- Check first. macOS:
ls /Applications/Always.app. Linux and Windows: always --version. If it is already installed, skip to step 5.
- Check the machine. macOS needs 14+ on Apple Silicon (
sw_vers -productVersion, and uname -m prints arm64). Linux needs glibc 2.39+ (ldd --version) on x86_64 or arm64. Windows needs 10 or 11 on x64. If the machine doesn't qualify, stop and tell the user why.
- Install with the one-liner for the OS from Install. On macOS and Windows it needs no admin rights; on Linux, installing the package may ask for
sudo.
- Finish setup.
- macOS: the installer opens Always. Tell the user the first-launch window asks for Microphone, Accessibility and (macOS 26+) Speech Recognition, and that only they can grant these.
- Linux and Windows: ask the user for their Groq API key (free at https://console.groq.com/keys) and run
always config set groq_api_key <key> — or, with no key, transcribe locally: always model download moonshine-medium-streaming-en && always model use moonshine-medium-streaming-en. Then start it: on Linux systemctl --user enable --now always-daemon.service; on Windows always run in a new terminal.
- Verify with
always status (on macOS: /Applications/Always.app/Contents/MacOS/always-daemon status). It must print Always-on daemon running. Report what it printed, and do not claim success otherwise.
- Tell the user they can start talking, and that on macOS Always only types into the apps they picked during setup (⌃⌥P adds the current one).
📚 Documentation
📝 License
AGPL-3.0. Use it, modify it, share it — if you distribute a modified version (including as a network service), publish your changes under the same license.
Built by humans who got tired of typing.
Website ·
Releases ·
Issues ·
⭐ Star on GitHub
