A single HTML page that watches you back via llama.cpp

What it does

This repository is essentially glue code: a single index.html page that captures your webcam stream and forwards frames to a locally running llama.cpp server loaded with SmolVLM-500M-Instruct. The model then describes what it sees, giving you browser-based vision inference that never leaves your machine. The README suggests you can tweak the prompt to get JSON output instead of prose.

The interesting bit

The entire frontend is a single static HTML file with no build step, no framework, and no package manager. That minimalism is the point: it demonstrates how little plumbing is needed to turn llama.cpp’s multimodal server into a real-time visual assistant.

Key highlights

• Runs entirely offline once the model is loaded locally
• Frontend is a single static HTML file with no dependencies
• Uses SmolVLM-500M, a compact vision-language model
• Supports custom prompts (e.g., requesting JSON responses)
• Works with llama.cpp’s standard server API

Caveats

• The README labels the output as “real-time object detection,” but it is unclear whether the model returns bounding boxes or just descriptive captions.
• Performance depends on local hardware; the README notes GPU offloading is likely necessary.
• Beyond the setup steps, there is essentially no documentation or error handling guidance.

Verdict

Worth a look if you want the shortest path from webcam to local vision model. Steer clear if you need documentation, error handling, or anything resembling a finished product.