An NVR that actually knows what it's looking at

What it does

Frigate ingests RTSP streams from IP cameras and applies on-device object detection using OpenCV and TensorFlow. It manages 24/7 recording with retention rules tied to detected events, serves low-latency live views over WebRTC or MSE, and pushes results into Home Assistant through a custom component and MQTT.

The interesting bit

Rather than grinding every frame through a neural network, Frigate uses cheap motion detection to decide exactly where and when object detection is worth running. It shards that work across separate processes and hands it off to a GPU or dedicated AI accelerator, keeping CPU load low enough to run on typical home servers.

Key highlights

• Entirely local inference; no cloud required for object detection
• Home Assistant integration via custom component and MQTT
• Motion-aware pipeline that only runs expensive detection where needed
• Multi-process TensorFlow architecture optimized for realtime FPS
• RTSP re-streaming to reduce camera connection load, plus WebRTC live view

Caveats

• The README treats a GPU or AI accelerator as nearly mandatory, noting that even fast CPUs lag behind dedicated hardware.
• The Frigate name and logo are trademarks of Frigate, Inc. and are explicitly excluded from the MIT license.

Verdict

Home Assistant users who want local AI object detection on their security cameras will find Frigate purpose-built for their stack. Those without a GPU or AI accelerator, and no interest in home automation, are not the target audience.