Replace Ring or Nest cameras with a local NVR

Check one thing before you read the rest of this page: does your camera expose an RTSP or ONVIF stream? If it does not, this is not a software migration. It is a hardware replacement with a software migration attached, and most Ring and Nest devices fall on that side. Their video path is designed to end at the vendor.

What you are actually giving up#

The subscription was not buying you storage. It was buying you somebody else's operations team, and here is the itemized list of what you take back.

The storage, and sizing it correctly. Recordings dominate everything else on a local NVR. Bitrate in Mbit/s times 10.8 gives GB per camera per day, which is just 86,400 seconds divided by 8 bits and scaled: a 4 Mbit/s stream is 0.5 MB/s, which is 43.2 GB a day. Six cameras for a week is roughly 1.8 TB. Nobody sends you a bill when you get this wrong. The disk simply fills.

The retention policy. Cloud plans hand you a number of days and enforce it. Locally you write the rule, and Frigate re-tiered retention in 0.17 with separate continuous and motion fields, continuous off by default. There is a safety net: under one hour of remaining space, Frigate deletes the oldest recordings regardless of your rules. That protects the host and silently discards footage you thought you had.

The false-alarm tuning. Cloud NVRs ship with someone else's tuned thresholds and a support queue. Locally you own the zones, the object filters, the minimum area, the required ratio and the score thresholds, and you will spend a fortnight being told about cats at 3am. Frigate's detection is genuinely good, which is the point of it over motion detection, but good still means you draw the zones.

The offsite copy. Cloud video was, incidentally, off-premises. Recordings that live in the same building as the camera are recordings an intruder can carry away. Almost nobody backs up NVR footage and that is defensible, but decide it on purpose. See Backups that actually restore: here the thing worth copying is /config, not the video.

Someone else's uptime. Frigate is still pre-1.0, every minor moves config keys, and database migrations are one-way. If the box is down there is no cloud recording of the hour you missed.

If that reads as work you do not want, keep the subscription. This is the highest-effort migration in the home category and it pays back for one specific objection: you do not want video of the inside and outside of your house held by a company that can hand it on or change its policy. Without that objection the value is thin.

Get your recordings out first#

Retention is tied to the subscription. When it ends, the history goes with it, and there is no export button that gives you the whole archive.

Do this before you cancel:

1. Download every clip you would be upset to lose, one at a time. There is no bulk export. Budget an evening.
2. Note the retention window you are actually on. That is the target the local system has to match, and it is usually shorter than people assume.
3. Inventory the hardware: model, RTSP or ONVIF support, whether it has a configurable second substream, PoE or Wi-Fi. That table decides your budget.
4. Test a stream before buying anything. ffprobe rtsp://user:pass@camera/stream gives you the resolution, codec and bitrate to plan around.

Step 3's substream question is not a detail. Detection runs at a fixed model input size regardless of what you feed it, so pointing the detect role at a 4K 20 Mbit/s main stream is the commonest reason a Frigate install pegs every core. Give detection a second stream at around 640x640 and a low frame rate, and reserve the high-resolution stream for recording. Fixing that on an existing install routinely turns 90 percent CPU into 15.

Which one to pick#

These two are not alternatives. Frigate is the NVR and the detector; Home Assistant turns a detection into a notification, a light and a rule about when not to bother you. Most people who finish this migration run both.

Run Frigate on a box you are not asking to do much else. An idle Intel N100 or an eighth-generation-or-later desktop is the sweet spot; Choosing home server hardware covers sizing.

What it actually costs#

Storage. Do the arithmetic before you buy the disk, not after.

The last row is a range because it depends on your street, and it is where the honest trade lives: detection-only recording is far cheaper and leaves you nothing when the detector misses. Two 4 TB drives in a mirror is the boring right answer for six cameras on continuous; ZFS, btrfs, mdadm or one disk covers whether to mirror at all.

The detector. A Coral TPU offloads inference at about 10 ms and does nothing for decode. A modern Intel iGPU on OpenVINO does inference in roughly 4 to 15 ms and also does the decode, which is the larger cost. The Frigate docs no longer recommend Coral for new installs outside low-power builds, and 0.17 dropped support for Nvidia GTX 900 series cards. Buy the iGPU, pass /dev/dri through, skip the USB stick.

Memory. Frigate idles around 700 MB and sits near 2.5 GB with a few cameras, because there are two ffmpeg processes per camera plus the detector. Semantic search changes the class of machine: the docs require 8 GB and recommend 16 GB.

Shared memory, which is not optional. Docker's default 64 MB is below what two cameras need. The formula is (width x height x 1.5 x 20 + 270480) / 1048576 plus 40 MB for logs, about 67 MB for one 1280x720 detect stream.

services:
  frigate:
    image: ghcr.io/blakeblackshear/frigate:stable
    shm_size: "512mb"
    devices:
      - /dev/dri/renderD128:/dev/dri/renderD128
    volumes:
      - ./config:/config
      - /srv/frigate:/media/frigate
      - type: tmpfs
        target: /tmp/cache
        tmpfs: { size: 1000000000 }
    ports:
      - "8971:8971"

Note what is missing: port 5000. Frigate exposes two HTTP ports deliberately. 8971 enforces login and role checks; 5000 is the internal admin API with no authentication at all, meant for containers on a trusted Docker network. Publishing or reverse-proxying it hands anyone your cameras and your config. For single sign-on, Frigate trusts X-Forwarded-User and X-Forwarded-Groups from Authelia or authentik; see Reverse proxy and TLS.

What breaks and how to tell early#

Notification latency, in week one. Walk up your own driveway with a stopwatch. The chain is camera, Frigate, detection, Home Assistant, push transport, phone. Detection is milliseconds; the seconds go to segment assembly and your automation. Over about ten seconds means a problem to find now, not the first time it matters.

False alarms, in week two. They come at night, from headlights, spiders on the lens and a flag. Draw zones and set object filters rather than raising the score threshold globally, because a threshold high enough to silence a moth silences a person in the rain too. The failure you are avoiding is someone in the household muting the notifications, and it is silent.

Cameras flapping offline. That is shm_size. Fix it before you look at the network.

The disk, in month two. Watch free space with Beszel or Uptime Kuma and alert on it. The safety net deleting your oldest footage is not an error you will see.

The point of no return#

Cancelling the subscription is not it. Running down the vendor retention window is: once the cloud history has aged out, the only copy of last month is the one on your disk, and if the disk was sized wrong or the retention config was wrong, you find out by needing footage that is not there.

So sequence it. Run Frigate in parallel with the subscription for one full retention cycle. Each week, look up one event in Frigate that you know happened and confirm the clip exists at the quality you would need. Only then cancel. And copy /config, meaning config.yml plus frigate.db, before every upgrade, because migrations do not go backwards.

What to do next#

Run ffprobe against one camera today; that answer decides whether this migration is possible at all. Then read the Frigate profile, size the disk from the table above and pair it with Home Assistant. If you are also leaving a cloud hub, Replacing SmartThings is the same weekend's work, and Home automation has the rest.