Frigate NVR

Frigate NVR is an MIT-licensed open-source network video recorder that turns IP cameras into a programmable source of AI events: all object detection runs on the local machine and the footage never leaves your LAN. Its integration with Home Assistant is deeper than most peers — a single MQTT message can drive an automation, rather than just offering another playback UI.

Core capabilities

• Local object detection: OpenCV + TensorFlow inference runs on the device; motion filtering comes first and detection follows, keeping resource use predictable
• Accelerator matrix: Coral, Hailo-8, Intel OpenVINO, NVIDIA, Apple Silicon and RKNN are all options; the official 100+ detections per second depends on a Coral-class accelerator plus a low-resolution substream, and CPU-only detection is recommended for testing only
• Face and license plate: built-in face recognition and license plate recognition write their results into event attributes and push them over MQTT
• Semantic search: local vector retrieval supports text-to-image and image-to-image search, gated on 8GB of RAM and a CPU with AVX2
• Generative summaries: connect OpenAI, Gemini, or local Ollama and llama.cpp; with the first pair the summaries go to the cloud, and only the latter pair keeps everything local
• Object-based retention: only recordings that contain a detected object are kept, and zones plus Birdseye make review fast

Who it suits People already running Home Assistant who want cameras to act as sensors; people with multi-brand ONVIF cameras to govern under one layer; tinkerers with hard privacy constraints who can accept running the machine fully offline. If you want appliance-grade plug-and-play, it is not for you.

Unique strengths Detection logic is fully programmable: zones, object filters, confidence and dwell time are tunable per camera, and both models and accelerator hardware can be swapped freely. Brand NVRs only give you rules hard-coded by the vendor, which is exactly why Home Assistant users accept the weekend of configuration work.

Editorial verdict The MIT core costs nothing, and face, license plate and semantic search are not behind a paywall; Frigate+ is only an optional model subscription. The price is hours to days of tuning, plus backup discipline before upgrades. It suits self-hosters who read the docs, not commercial deployments that expect the vendor to carry the risk.