A desktop pet that actually watches your screen and thinks

What it does

B is a PyQt6 overlay creature that lives on your desktop: a 60fps animated face that tracks which window you’re focused on, extracts its content semantically (via UI Automation tree walking, with a Tesseract OCR fallback for apps that don’t cooperate), and feeds that context to an LLM. It listens via Faster-Whisper, talks back through Piper ONNX TTS with a DSP chain (pitch shift, bitcrush, chorus — deliberately robotic), and can decide on its own when to interject.

The interesting bit

The architecture is the real story: every module talks exclusively through a thread-safe pub/sub EventBus, so the window tracker, vision sensors, LLM brain, and face renderer never know about each other. That discipline is unusual in a hobby-scale companion app, and it’s what makes the perception → cognition → expression pipeline coherent rather than spaghetti.

Key highlights

• Semantic screen reading via UIA tree walks, with quality scoring, adaptive cooldown, and OCR fallback
• Dual LLM backends: Groq cloud or local llama-cpp (GGUF models, gitignored) for private inference
• An AutonomyEngine that schedules unprompted “proactive thoughts” based on context quality and timing
• Emotional state machine (brain/soul.py) driving blinking, resting, and conversing expressions
• Physics-based movement with Bezier interpolation; a Work Mode (Ctrl+Shift+Alt+W) shifts it toward utility

Caveats

• The README leans heavily on vibe (“soul-wired digital lifeform”) and is truncated mid-section, so the actual capabilities of Work Mode beyond “semantic monitoring” are unclear
• Windows-centric: UI Automation and MSS capture mean this won’t run meaningfully elsewhere
• “Silent by default” is a design claim — how well the autonomy loop respects it in practice isn’t documented

Verdict

Worth a look if you’re building desktop agents, screen-context pipelines, or just want a well-decoupled event-bus architecture to study. If you wanted a plain command-driven assistant, this is deliberately not that.