Type and speak with your eyebrows and your breath.
Download · How it works · Privacy · Run from source
A webcam switch-access keyboard for people who can't reliably use their hands or voice, built on the Presage SmartSpectra Node SDK. No special hardware, no cloud service, no account.
A highlight steps through the board on its own. You need two gestures:
| Gesture | Does | Read from |
|---|---|---|
| Raise your eyebrows | Select the highlighted row, then the highlighted key | face landmarks (brow-to-eye distance) |
| Breathe out slowly (about 3 s) | Speak the sentence aloud and copy it to the clipboard | chest breathing trace |
- Back and delete are keys:
↩ends every row and⌫is on the first letter row. - Letters are in frequency order (space, E, T, A, O first), so the common ones take the fewest steps.
- Quick phrases on the top row ("Yes", "I need water", "I need help" …) are spoken as soon as you select them.
- Send to chat (optional): set WhatsApp, Telegram or Email in Settings and a Send key appears on the last letter row. It opens that app with your text filled in; you still press Send there.
- A keyboard or physical switch works too:
Spaceselect ·Enterspeak ·Escstop.
| SDK signal | What Glint does with it |
|---|---|
face.landmarks (478 points) |
Brow-to-eye distance ÷ eye spacing → select when it rises about 28% and holds |
breathing.upperTrace |
Rises on inhale, falls on exhale. A gentle fall of ≥ 6 units for ≥ 2.5 s → speak |
face.talking |
Blocks speaking while you are talking |
validationStatus + hint |
Scanning pauses and shows the SDK's own hint ("Increase light on face", or a hint when more than one face is visible …) |
cardio.pulseRate |
Shown in the strip so you can see you're being read. Never used for any decision |
The detectors are small pure functions (src/renderer/gestures/), tuned and tested on a real recorded session with timestamped gesture cues (recordings/guided.jsonl: numbers only, no video). Everything the SDK measured is in docs/sdk-notes.md.
Why eyebrows and not blinks? The plan was a long blink to select. On the recorded session, five deliberate 1-second eye closures were invisible to both the SDK's blink signal and the face landmarks: the user wears thick-framed glasses and sat in dim light. Eyebrow raises showed up every time (+40%). So the gesture set follows the measurements.
- The camera starts only from the consent screen and is released on Stop, Esc, any error that ends the session, or closing the app.
- Video is processed on the device by the SDK. Glint never writes it to disk or sends it anywhere.
- Nothing you type or say is stored. Spoken text goes to the clipboard.
- Text leaves the computer only if you set Send to in Settings and select the Send key. Glint then opens WhatsApp, Telegram or your mail app with the text filled in, and you still press Send there. The recipient is kept in
settings.json. - The only stored file is
settings.jsonin the app's user-data folder. Forget everything deletes it. - SDK telemetry (session summaries and readiness reports; no frames, no vitals, but not anonymous: it can be linked to your Presage account) is off unless you turn it on. See Presage telemetry & privacy.
- No LLM Insights, no AI service, no analytics, no accounts.
- Glint is not a medical device and makes no health judgements.
Needs Node ≥ 20 and a free API key from physiology.presagetech.com.
npm install
cp .env.example .env # put your key in PRESAGE_API_KEY
npm start # GLINT_DEVTOOLS=1 npm start opens DevToolsNo key or no camera? Choose Try a recorded session on the first screen. It replays a recorded session through the same detectors, and you can type with Space / Enter.
Linux speech: most distros give Chromium no built-in voice, so Glint falls back to espeak-ng (sudo pacman -S espeak-ng / sudo apt install espeak-ng). Without it, text is still copied.
Light matters. Sit facing a lamp or window. A face in shadow is the most common reason tracking fails.
The SDK supports Linux x64 (glibc ≥ 2.35), Windows x64 and macOS arm64.
Builds are on the Releases page (made by CI): Glint-<version>-linux-x86_64.AppImage and Glint-<version>-win-x64.exe. They are unsigned, so Windows SmartScreen will ask you to confirm.
- Linux:
chmod +x Glint-*.AppImage && ./Glint-*.AppImage. If it says AppImages require FUSE (Arch, Ubuntu 22.04+), runAPPIMAGE_EXTRACT_AND_RUN=1 ./Glint-*.AppImageor installlibfuse2. - API key for an installed app: set
PRESAGE_API_KEYin your environment, or putPRESAGE_API_KEY=…in a.envfile in the app's user-data folder (~/.config/Glint/on Linux,%APPDATA%\Glint\on Windows). The key is never bundled into a build.
Build locally with npm run dist:linux (or dist:win on Windows). Output goes to out/.
yashksaini-coder.github.io/glint. The source is in site/, built with Fumadocs: cd site && npm install && npm run dev. Pages are plain MDX in site/content/docs, starting with Scope (what has and hasn't been tested).
npm test # detectors on the recorded fixture, scanner, settings validation
npm run smoke # launches the app, replays, types + speaks via simulated switch presses
node tools/screenshots.ts # regenerates the app screenshots in docs/img/ (replay mode: no camera, no face)
npm run diagram # regenerates docs/img/architecture.png from tools/diagram/make.mjs
cd site && npm run og # regenerates the share images (site/public/og.png, docs/img/social-preview.png)
node tools/sweep.ts # grid-searches detector thresholds against the recorded cuessrc/main/ Electron main: SDK IPC, camera permission, settings file, clipboard, espeak fallback, send links
src/preload/ bridge (SDK preload + a small window.glint API)
src/renderer/ app.ts (consent → setup → board), gestures/, board/, signals/ (live + replay)
src/shared/ settings schema + sanitizer, send link builder (used on both sides)
recordings/ guided.jsonl test fixture
tools/ analysis, tuning, smoke test, screenshots
Entry for the DEV × Presage SmartSpectra challenge (Nov 2026). Done: consent flow, scanning board, quick phrases, settings, send to chat (WhatsApp, Telegram, email; you still press Send), both detectors, replay mode, Linux/Windows builds. Next: per-user calibration, demo video.
Glint's own code is MIT. The Presage SmartSpectra SDK (@smartspectra/node-sdk and its platform runtimes) is proprietary and under Presage's own terms; packaged builds include its runtime and its bundled third-party notices (resources/smartspectra/).




