TL;DR:

  • Frigate is a free, open-source NVR that runs AI object detection locally — no cloud, no subscription, video stays on your network
  • Hardware baseline in 2026: Raspberry Pi 5 (4GB+) or an Intel N100 mini-PC handles 2–4 cameras on CPU; add a Coral TPU or Raspberry Pi AI Kit (Hailo-8L) to handle 6–10+ streams at 10% CPU load
  • Hailo-8L (13 TOPS, Raspberry Pi AI Kit) now beats Coral Edge TPU (4 TOPS) on price-per-TOPS for most builds; Coral still works well for ultra-budget 1–3 camera setups
  • Frigate integrates natively with Home Assistant for automations triggered by detected objects (person, car, dog)
  • Detection latency drops from 80–200ms on CPU to 6–15ms with a TPU — meaningful for real-time alert triggering

Cloud-connected security cameras have a particular failure mode: the company hosting your footage shuts down, raises prices, or gets acquired and changes the terms. Frigate exists as an answer to that — an open-source NVR that does AI object detection on hardware you own, with footage that never leaves your network.

In 2026, the hardware equation for running Frigate has improved significantly. Dedicated edge AI accelerators are cheap, the Raspberry Pi AI Kit is widely available, and the Intel N100 mini-PC has become the de facto affordable server for home lab edge workloads. The software has matured too: Frigate 0.14 added improved model support and a cleaner configuration interface. This is now a genuinely practical system for anyone comfortable with a bit of YAML and Docker.

What Frigate Does

Frigate connects to IP cameras via RTSP streams, runs real-time object detection on each stream, stores detected event clips, and provides a web interface for reviewing footage and configuring detection zones.

The detection output is a stream of events: “person detected in zone: front_garden at 14:32:07” with a 30-second clip attached. These events are published to MQTT, which is where the integration with Home Assistant and other home automation systems happens. A detected person at the front door can trigger a notification, unlock a smart lock, turn on lights, or do whatever your automation platform supports.

Critically, this all runs locally. No video is uploaded anywhere. Object detection runs on your hardware, triggered by motion in the configured zones, with clips stored on local disks.

Hardware Options in 2026

CPU-only (Raspberry Pi 5 or Intel N100): A Pi 5 with 4GB RAM handles two cameras at 5 frames per second for detection without additional hardware. An Intel N100 mini-PC — typically under £150 — handles four to six cameras comfortably. CPU detection latency is 80–200ms, which is fine for recording and notification but adds perceptible delay to real-time automations.

Coral Edge TPU (4 TOPS): Google’s USB Coral accelerator drops detection latency to 6–15ms and offloads inference entirely from the CPU. It connects via USB 3.0 and is supported natively in Frigate’s configuration. At £70–£90, it is still a reasonable addition for 1–3 camera builds where the Pi 5’s CPU would otherwise be saturated.

Hailo-8L via Raspberry Pi AI Kit (13 TOPS): The Raspberry Pi AI Kit connects via the PCIe interface built into the Pi 5, giving 13 TOPS of inference capacity. This handles 4–5 simultaneous camera streams in real time with the Pi 5’s CPU sitting around 10% load. At current pricing it beats Coral on cost-per-TOPS and is the recommended accelerator for Pi 5-based builds as of mid-2026.

Hailo-8 (26 TOPS) on mini-PC: For larger installations — 8+ cameras — a mini-PC with a Hailo-8 M.2 accelerator handles the load without breaking a sweat. This is overkill for most home deployments but appropriate for small commercial premises.

Basic Configuration

Frigate runs as a Docker container. A minimal docker-compose.yml:

services:
  frigate:
    image: ghcr.io/blakeblackshear/frigate:stable
    privileged: true
    restart: unless-stopped
    ports:
      - "5000:5000"
      - "8554:8554"
    volumes:
      - /path/to/config:/config
      - /path/to/storage:/media/frigate
    environment:
      FRIGATE_RTSP_PASSWORD: "your_camera_password"

A minimal config.yml:

cameras:
  front_door:
    ffmpeg:
      inputs:
        - path: rtsp://admin:{FRIGATE_RTSP_PASSWORD}@192.168.1.50:554/stream1
          roles:
            - detect
            - record
    detect:
      width: 1280
      height: 720
      fps: 5
    objects:
      track:
        - person
        - car

detectors:
  coral:
    type: edgetpu
    device: usb

Home Assistant Integration

Frigate has a first-class Home Assistant integration available through HACS (Home Assistant Community Store). Once installed, every camera and detected object type becomes a sensor in Home Assistant. You can build automations like:

  • Notify phone when a person is detected at the front door between 22:00 and 07:00
  • Turn on floodlight when a car enters the driveway
  • Log all animal detections to a history file

The integration also adds Frigate camera feeds directly to the Home Assistant dashboard, with clip playback and event timeline built in.

What to Watch Out For

Camera compatibility: Frigate works with any camera that provides an RTSP stream. Most IP cameras do. Some consumer cameras lock to their manufacturer’s cloud platform and do not expose RTSP. Check before buying — Reolink, Amcrest, and Hikvision cameras are commonly recommended for Frigate compatibility.

Storage: Detected event clips accumulate. For a four-camera system detecting regularly throughout the day, expect 20–50GB per week depending on clip retention settings. Configure retain settings in Frigate to limit storage consumption.

False positives: Default YOLOv8-based detection models will occasionally flag shadows, rustling foliage, or reflections as motion. Mask regions of the camera frame that generate consistent false positives, and use Frigate’s min_score threshold settings to reduce noise.

scale_from_zero: If you have cameras that rarely trigger detection, Frigate can scale detector usage down when idle. For latency-sensitive setups, disable this or set minimum detector allocation to avoid the warm-up delay on first detection.

Frigate represents the edge computing AI inference pattern at its most practical: commodity hardware, an open model, local processing, real output. The cloud subscription model for home security is not the only option.