Frigate: a 37k-star open NVR that detects people, cars and pets locally

The default home camera sends your footage to the vendor’s cloud — privacy, monthly fees, and blindness when the internet drops. Frigate (MIT, maintained since 2019, 37,268 stars as of November 5, 2026) is the opposite design: a complete local NVR that runs realtime object detection on IP camera feeds with OpenCV and TensorFlow on your own hardware — people, cars and pets, no vendor cloud involved. The WhatsApp note that surfaced it got the shape right: events flow over MQTT into Home Assistant, WebRTC gives low-latency multi-camera live view, and the whole thing is performance-minded.

Core features

• Local realtime detection: TensorFlow runs in separate processes for maximum FPS, and a low-overhead motion detector decides where detection is worth spending cycles.
• Deep Home Assistant integration: an official custom component plus MQTT, so detection events drive automations directly.
• 24/7 recording: retention settings keyed to detected objects; RTSP restreaming reduces connections to each camera.
• Low-latency viewing: WebRTC and MSE both supported, multi-camera grid included.
• Accelerator friendly: a maintained list of supported object detectors; a GPU or AI accelerator (Coral and friends) far outperforms CPU-only.

Typical use cases

• Self-hosted home security: person detection at the door, pet activity logs, events wired into Home Assistant automations.
• Small shops: a few cameras retained locally, nothing passing through a third party.
• Detection-rule tuning: masks and zone editors make false-positive cleanup visual.

Quick start

Docker deployment is the main path (full compose configs in the docs), cameras come in over RTSP; budget for at least one AI accelerator — CPU-only works but limits camera count. The detector ecosystem is mature: Google Coral TPU, OpenVINO and ONNX all have official support paths, so an old machine with a Coral stick drives several cameras comfortably.

Summary

Frigate fits self-hosted home and small-business security where footage never leaves the premises; teams wanting vendor-style cloud playback or face databases should compare commercial NVRs. Deployment is Docker-first, and Home Assistant users get an official integration component — detection events appear as cards on HA dashboards without glue code, and automation rules use stock triggers. Facts and limits: MIT, Python, maintained since 2019; detection models need per-scene tuning, false-positive control lives in masks and zones, and recording storage scales with camera count and retention — budget the disk for a rolling window. It belongs to the “your own cloud at home” series on this site — alongside agent-autoloop and auto-ontology: perception and decision-making that used to be handed over now stay on your machine.