Transparent screen-overlay teleprompter for Linux/Win/Mac. Multi-file playlist, voice-paced auto-scroll, screen-capture hiding, cue points, markdown + docx ingest. Built on Electron + React + TypeScript.

npm install
npm run dev

Two windows open:

• Overlay — frameless, transparent, always-on-top. Hover near the top edge to reveal a drag handle. Double-click the text to jump straight into the editor.
• Controls — playlist, sliders, toggles, hotkey reference, settings.

• Multi-file playlist (.txt, .md, .rtf, .docx) via file picker or drag-and-drop; recent paths persist
• Variable opacity (5–100%) + background dim
• Auto-scroll, 5–400 px/s
• Typography: font size, family, color, drop-shadow
• Markdown rendering (sanitized via DOMPurify) — headings, lists, blockquotes, inline code
• Eye-line marker + focus mode (masks lines outside the eye-line band)
• Mirror modes: horizontal (beam-splitter rigs) and vertical
• Banner / lower-third mode — single-line horizontal ticker at top or bottom
• RTL — dir="auto" on text containers; per-paragraph direction detection
• Chronometer in overlay corner: elapsed · time-to-end · target WPM
• 3-2-1 countdown before play (configurable seconds, only when starting from the top)
• Cue points — [[CUE: name]] markers become a clickable list and bind to Ctrl+Alt+1..9; optional in-overlay cue HUD shows the upcoming list with the current cue highlighted

• Live edit pane — debounced (200 ms) writes to in-memory state; save back to disk (or save-as for new scripts)
• Edit-while-prompting — double-click overlay text to open the editor and focus the controls window

• Click-through mode — mouse passes through to apps below
• Hide from screen capture — setContentProtection (macOS / Windows; Linux unsupported and indicated in UI)
• Per-window geometry persistence; bounds clamped to current displays on launch

• Global hotkeys (see table below); rebindable per-command from the Hotkeys panel
• Clicker mode — registers PageUp / PageDown globally to step scroll by a configurable amount; opt-in toggle
• Voice pacing — Web Speech API matches your spoken words against the script and auto-advances. Requires consent on first enable (styled modal); status pill reads "voice (cloud)" while active because Chromium routes audio to Google for transcription

• Reset to defaults (with confirm)
• Export / import config (JSON)
• About: app, electron, node versions; on-disk store path

Default bindings:

Each binding (except cue jumps and clicker keys) is rebindable in the Hotkeys panel — click an accelerator to capture a new key combo, or hit ↺ next to it to restore default. Failed registrations (held by another app or the compositor) are listed in red.

npm run build       # transpile main/preload/renderer
npm run package     # build + electron-builder distributables
npm run typecheck

src/
├── main/         Electron main process — windows, IPC, hotkeys, electron-store state
├── preload/      contextBridge → window.api
├── shared/       AppState, cue parser (used by main + both renderers)
└── renderer/
    ├── overlay.html  + src/overlay/    Transparent scrolling window
    ├── controls.html + src/controls/   Control panel
    └── src/shared/voice.ts             Web Speech API + sliding-window alignment

• State lives in the main process (electron-store with debounced 250 ms disk writes); broadcast to both renderers on every patch.
• Persistence stores file paths only — file contents are re-read on launch and never touch disk via electron-store.
• Auto-scroll is driven by a single RAF loop in the overlay with a liveRef snapshot of state and 50 ms-throttled IPC dispatch back to main; echo-suppression prevents the writer from reprocessing its own update.

• contextIsolation: true, nodeIntegration: false, sandbox: true on both BrowserWindows
• setWindowOpenHandler({ action: 'deny' }) and will-navigate blocking external URLs
• Strict Content-Security-Policy (per-directive, no 'unsafe-inline' on scripts)
• DOMPurify sanitizes markdown HTML before injection
• IPC handlers validate every input against an explicit PATCHABLE_KEYS allowlist with type checks and range clamping
• files:loadPath rejects paths outside the recent set, session-allowed set, or whitelisted extensions
• files:save to non-session-allowed paths is forced through a save dialog
• Permission request handler allows only media (mic for voice pacing); all others denied
• Path-only persistence keeps file contents off disk via electron-store

• Wayland: alwaysOnTop, global hotkeys, and screen-capture hiding behave inconsistently across compositors. The Controls window shows a banner on Wayland; XWayland is the most predictable fallback.
• Linux generally: setContentProtection is a no-op on Linux (Electron limitation). The toggle is disabled with an explanation.
• Voice pacing: webkitSpeechRecognition in Electron 41 routes audio to Google for transcription. First-enable shows a consent dialog. The status pill reads voice (cloud) while active.

• RTF parsing is hand-rolled; complex RTF (embedded objects, tables) may strip imperfectly. Use .docx for higher-fidelity source.
• Voice pacing is cloud STT (Chromium routes audio to Google). No local fallback yet.
• Ctrl+Alt+1..9 (cue jumps) and PageUp / PageDown (clicker) are not rebindable in v1.