Self-hosted AI object detection server for BlueIris NVR — a drop-in CodeProject.AI replacement.
Point BlueIris at it and it just works — no reconfiguration, no cloud. Supports 30+ models across YOLO v5–v26, RT-DETR, Faster R-CNN, SSD, and RetinaNet. Runs on CPU, NVIDIA CUDA, Intel OpenVINO, or AMD ROCm.
- Install Python 3.10+ — tick Add to PATH
- Download the latest release and extract it
- Run
start.bat - Open http://localhost:32168, pick a model from Model Browser, click Download → Load
- In BlueIris → AI → Configure, set the Server URL to
http://<this-machine-ip>:32168
For auto-start on login: run setup-service.bat. Full docs in the dashboard Help tab or USER_GUIDE.md.
- Drop-in BlueIris replacement — same API as CodeProject.AI, zero reconfiguration
- 30+ detection models — switch between them in the dashboard
- One-click GPU setup — install CUDA or OpenVINO from the Hardware tab
- Legacy GPU support — older NVIDIA cards (GTX 10-series, Tesla P4) work via ONNX Runtime
- Live console with per-detection confidence, 11 themes, optional auto-start
LAN-only by design. Do not expose port 32168 to the internet. Built-in mitigations include a Host-header allowlist (blocks DNS rebinding), per-IP rate limiting on the detection endpoint, and single-use WebSocket tickets. See USER_GUIDE.md for details.
This project was built collaboratively with AI (Anthropic's Claude). The code and documentation were AI-generated; I directed the design and tested against real BlueIris deployments.
AGPL-3.0 — uses Ultralytics YOLO, which is also AGPL-3.0. Any redistribution must include source code. Commercial use requires an Ultralytics Enterprise License.
| Version | Notes |
|---|---|
| v0.7.8 | Security hardening — Host-header allowlist, rate limit, WS tickets, config validation |
| v0.7.7 | Removed system tray. Scheduled-task path fixes |
| v0.7.6 | YOLOv5 switched to direct GitHub release URLs |
Earlier versions documented in HANDOFF.md.