A native macOS AI interview assistant (codebase target still named Smarty). Cueglass listens to your mic, optionally reads the screen via OCR, and generates spoken-style answers into a floating overlay with best-effort screen-sharing exclusion.
No accounts, no backend, no analytics. Pick an AI provider in Settings, paste its API key (Keychain only), and go.
- Multiple AI providers — OpenAI, Claude (Anthropic), Gemini (Google), or DeepSeek for answers; switch anytime in Settings → API. Each provider keeps its own key and model list, so switching doesn't lose the others.
- Whisper / Stream modes — tap-to-record + Send, or Listen with pause detection
- Floating overlay with best-effort blind mode (
sharingType = .none) - Region capture and answer — press Control–Option–S, drag over a question, and release to send the cropped screenshot; Escape cancels
- Screen OCR and typed ask with optional image attachments (multimodal on OpenAI, Claude, and Gemini; DeepSeek is text-only)
- Role brief — set the job title, field, company, and notes for the role you're interviewing for; answers are framed for that field (non-technical roles drop code, algorithms, and Big-O)
- Session presets — interview focus, answer length, preferred coding language
- Key terms panel and Markdown transcript export
- Global shortcuts for overlay, pause, and start
- Apple Silicon Mac (recommended)
- macOS 14.0+
- Xcode 16+
- An API key for at least one provider: OpenAI, Anthropic, Google AI Studio (Gemini), or DeepSeek
- For voice input (Whisper/Stream modes): an OpenAI key, or a Gemini key with Gemini selected as the answer provider — Claude and DeepSeek have no speech-to-text
-
Open or regenerate the project:
brew install xcodegen # once xcodegen generate open Smarty.xcodeproj -
Select the Smarty scheme → My Mac.
-
Signing (required for Screen Recording):
- Xcode → Signing & Capabilities → your Team (Apple Development).
- Avoid ad-hoc / “Sign to Run Locally” — TCC will not stick across rebuilds.
-
Run (⌘R).
-
Settings → API → pick a provider (OpenAI, Claude, Gemini, or DeepSeek), paste its API key → Save. Each provider's key is stored in Keychain only, under its own entry — never commit
.envor key files (see.gitignore). Voice input (Whisper/Stream) uses Gemini when it's selected and has a saved key, otherwise it falls back to OpenAI. -
Grant Microphone / Speech / Screen Recording when prompted.
- Screenshots and attachment frames are processed in memory (not written as a capture archive)
- Each provider's API key lives in its own Keychain entry
- Settings, history, and window frame use UserDefaults
- Only traffic to your selected AI provider (and OpenAI, unless Gemini is handling voice input) leaves the machine
- Region screenshots are sent to your selected provider for visual answering (OpenAI, Claude, and Gemini support this; DeepSeek is text-only), with optional OCR; typed drafts and pending attachments remain separate
- No authentication / cloud account for Cueglass itself
Blind mode requests exclusion for the selection panels and answer overlay before they appear.
Apple documents NSWindow.SharingType
as a legacy setting: recording apps may still capture these windows. This feature cannot guarantee
invisibility in Google Meet, Zoom, Teams, TestGorilla, or other recording/assessment platforms.
Cursor movements, macOS permission indicators, the menu bar, and changes in other apps may remain visible.
Verify the remote viewer’s output for your exact app, browser, macOS version, and sharing mode.
| Shortcut | Action |
|---|---|
| ⌃⌥S | Capture region and answer — global; drag then release, Escape cancels |
| ⌘⇧S | Start session (Stream) — app must be frontmost |
| ⌘⇧H | Show / hide overlay — global |
| ⌘⇧P | Pause / resume — global |
Region capture works without starting a listening session. Configure the API key and grant Screen Recording access using Settings → General → Permissions before using it; the shortcut does not open permission prompts. Select an area at least 8 × 8 points on one display. A drag crossing displays is clipped to its starting display. Wait for an existing answer to finish, or resume a paused session, before selecting another question. If another app owns ⌃⌥S, use Capture Region and Answer in the menu bar, or free the shortcut and restart.
Cueglass opens directly into its floating overlay, with no separate main window. Closing the overlay hides it while the app stays running. Press ⌘⇧H or reopen the app to bring the same overlay back.
Cueglass has no Dock icon (LSUIElement). The menu bar item (the <Cue/> wordmark) toggles the
overlay, opens Settings, starts/stops a session, and quits. Hide it under Settings → General → Menu Bar;
the shortcut and reopening the app still work when it is hidden. Blind mode does not hide the menu bar itself.
The overlay’s gear icon and menu bar open Settings on demand. Open Settings → General → Permissions for microphone and screen-recording setup; Settings and Permissions never open automatically at startup.
App/ SwiftUI entry + DI (AppEnvironment)
Views/ Permissions, ask field, role editing, history
Overlay/ Non-activating NSPanel + glass UI
Settings/ Provider/API, capture, overlay, prompt presets, general
Managers/ InterviewSessionManager, OverlayManager
ScreenCapture/ ScreenCaptureKit frames (in-memory)
OCR/ Vision text extraction
Speech/ Mic capture + OpenAI/Gemini STT / pause detection
OpenAI/ AIProviderClienting + SpeechTranscribing protocols; OpenAI Responses API client (text + multimodal + STT)
AIProviders/ AIClientRouter, plus Claude / Gemini (also does STT) / DeepSeek clients
PromptBuilder/ Context store, summarization, prompt assembly
Services/ Keychain (per-provider keys), settings, permissions, hotkeys, login item
Models/ Shared types (settings, AIProvider, attachments, messages)
Answer generation is provider-agnostic: InterviewSessionManager and ContextSummarizer depend on
AIProviderClienting, and AIClientRouter dispatches each request to the client for the currently
selected AIProvider. Speech-to-text goes through OpenAIClient or GeminiClient directly
(both implement SpeechTranscribing), selected per-session based on the answer provider;
Claude and DeepSeek offer no transcription API today.
MVVM: views bind to @Observable managers; capture / OCR / OpenAI run in actors.
See .env.example. No provider API key should ever be committed. Enter each key in Settings → API so it is stored in Keychain (one entry per provider). .gitignore already excludes .env, key files, and Xcode user state (xcuserdata/, *.xcuserstate).
xcodegen generate
xcodebuild -scheme Smarty -destination 'platform=macOS' -configuration Debug build
xcodebuild test -scheme Smarty -destination 'platform=macOS'