Build a Frigate camera box

Frigate is a self-hosted NVR - network video recorder - that watches your cameras, uses AI to tell people and cars apart from waving trees, and keeps everything on your own hardware with no monthly fee and no footage in someone else's cloud. It is one of the most satisfying home-server projects because it replaces a subscription with a box you own. Here is how to build one that runs smoothly.

1. The AI accelerator is the key part

Frigate runs object detection on every camera's video. Doing that on a CPU alone will bury even a strong machine once you add cameras. The fix is a cheap dedicated accelerator that does the detection work:

• A Google Coral (USB or M.2) is the classic, low-power choice and runs Frigate's detector beautifully for a handful of cameras.
• An Intel iGPU with OpenVINO, or a Hailo module, are strong modern alternatives, and the iGPU may already be in your box.
• A discrete GPU is overkill for detection alone but makes sense if the same machine also transcodes.

Which accelerator suits how many cameras - and which are actually buyable - is exactly what ServerRigs compares. Sort that out before anything else, because it sets the whole build.

2. Pick the box

Frigate runs happily on a mini PC with a modern Intel iGPU and a free USB or M.2 slot for the accelerator. That combination handles several cameras at low power and no noise. Give it enough RAM for the OS, Frigate, and its detection buffers, and put it where it can stay on all the time. If you already run a homelab, Frigate is a natural container to add to it.

3. Choose cameras that behave locally

Frigate wants a clean, standard video stream, which means cameras that speak RTSP and support two streams - a high-resolution one to record and a low-resolution one for detection. Avoid cameras that only work through a vendor cloud app; they fight you at every step.

Prefer PoE (Power over Ethernet) cameras. One cable carries both power and data, the feed is reliable, and nothing depends on Wi-Fi holding up in a storm. What to look for in a camera - sensor, low-light behaviour, lens - is covered at CameraScore.

4. Build the network to match

PoE cameras need a PoE switch to power them, and camera traffic is heavy and constant, so give it room. Two habits worth adopting:

• Put the cameras on their own VLAN so a cheap camera's firmware can never reach the rest of your network or phone home unsupervised.
• Size the PoE budget of the switch to the number and wattage of cameras, with headroom.

The PoE switches and the network gear to isolate cameras cleanly are at HomeSwitches. Locking cameras to a VLAN with no internet access is the single biggest privacy and security win in the whole build.

5. Size the storage for continuous recording

Cameras recording around the clock generate a lot of data, and it never stops. Two points:

• Use drives rated for surveillance or NAS duty, which are built for constant writes from many streams. A desktop drive is the wrong tool. Which drives suit continuous video - and which are CMR - is at NASdisks.
• Decide your retention (how many days you keep) and size capacity from the camera count, resolution, and that window. Frigate lets you keep detected events longer than the raw continuous footage, which stretches the disk a long way.

Keep the footage on storage you can grow, ideally a separate NAS, so adding cameras later is a capacity decision, not a rebuild.

6. Keep it private and backed up

The entire appeal of Frigate is that the footage is yours. Do not undo that by exposing it to the internet. Reach it from outside with a VPN, not a port forward. Back up Frigate's configuration - the part you would hate to recreate - to your NAS, and you are done.

The short version

Choose the AI accelerator first, run Frigate on a quiet mini PC, use RTSP PoE cameras on their own isolated VLAN, power them from a PoE switch with headroom, and record to surveillance-rated storage sized for your retention window. The result is a private, no-subscription NVR you fully control. Start with the accelerator comparison at ServerRigs and the cameras at CameraScore.