Your CCTV camera, now with an opinion

What it does DeepCamera is an open-source AI skills platform for security cameras. It ingests RTSP or ONVIF feeds, runs object detection and VLM scene analysis entirely offline using models like Qwen, DeepSeek, and YOLO26, then chats with you about what it saw through Discord, Telegram, or Slack. Think of it as a Home Assistant for paranoid people who own GPUs.

The interesting bit The architecture is deliberately modular: every capability is a “skill” — a self-contained module speaking JSONL over stdin/stdout. The clever part is an LLM-powered deployment agent that reads each skill’s SKILL.md manifest, probes your hardware (NVIDIA, Apple Silicon, Intel, AMD, or Coral USB), installs the right dependencies, converts models to the optimal format, and verifies it all works. No bespoke installers; the LLM does the yak shaving.

Key highlights

• Hardware auto-detection via shared env_config.py — TensorRT for NVIDIA, CoreML for Apple Silicon, OpenVINO for Intel, ONNX for AMD/CPU
• YOLO26 detection (Jan 2026 revision, no NMS/DFL) with claimed ~2ms–25ms latency depending on model size
• 143-test home-security-benchmark suite for evaluating VLM security performance on your own machine
• Skills include depth-map privacy transforms, SAM2 segmentation, dataset annotation, and agent-driven YOLO fine-tuning
• Companion desktop app “Aegis” handles one-click setup, camera connection, and conversational agent interface

Caveats

• Several headline skills (face recognition, license plate recognition, model training, Home Assistant bridge) are marked “planned” (📐) not ready
• The project is primarily JavaScript per GitHub metadata, though the ML stack and skills are Python-heavy; the repo language classification may mislead
• Some benchmark latency claims lack reproducible methodology in the README

Verdict Worth a look if you want local AI surveillance without cloud subscriptions and enjoy tinkering with model backends. Skip it if you need turnkey face recognition today or run purely on Raspberry Pi without acceleration.