How Helen Gets Made A making-of, not the live show Watch her live

Frigate NVR · setup we actually use

How to Set Up Frigate NVR for a 24/7 Cat Cam

Let Frigate watch the room so you do not have to sit there.

This is how we run Frigate NVR on a 24/7 cat cam. Install it, add one camera per job, turn on cat detection, and treat each hit as a visit you can cut.

Helen, a dilute tortoiseshell calico, already in frame — the kind of visit Frigate is for
A visit starts when she is in the picture. If frame one is an empty room, you waited too long or trusted the wrong cue.

Frigate is an open-source camera watcher people use at home. It looks at the cameras you already have, notices when something it was asked to find walks into the picture, and keeps a short record of that moment. We ask it to find a cat. It runs on a regular Windows PC that stays on. No graphics card, no Coral stick: since August 2026 the finding is done by the Intel graphics built into the processor itself, through OpenVINO, at about 12 milliseconds a look. The project lives at frigate.video. If you want the one-paragraph version of what Frigate is before the steps, read what Frigate NVR actually does all day.

Pick your page — the Frigate cluster
The questionThe page that answers only that
What is Frigate, in plain words?What is Frigate NVR? Detect stream, record stream, detector, retention: the three ideas, one sentence each.
Can I run it on Windows?Frigate on Windows with WSL2 and Docker, the way we run it 24/7 on a desktop PC. The exact files are on the setup page.
How do I add a camera?Camera setup: RTSP, sub-streams, go2rtc, with a Tapo example.
Can I see your whole config?The full config.yml for eleven cameras, annotated line by line.
Do I need a Coral?Probably not. CPU, Intel iGPU, Coral and NVIDIA compared with what we measured.
How does the iGPU do it under WSL2?An OpenVINO detector over ZMQ: the full server, unit file and numbers.
How much disk will this eat?Retention and storage math from a real 3.7 TB drive.
How do I put a camera on a public site safely?The public cam proxy: four cameras public, Frigate never on the internet.
How do the visits become Shorts?The Shorts pipeline, and the two 24/7 streams that never stop.
The computer, with no graphics card
Ordinary partWhat it does
A regular Windows PC that stays onWatches the streams while we sleep. No one sits at a desk calling action. How Frigate runs on Windows.
No graphics card. No Coral.Frigate’s cat detection runs on the Intel graphics inside the processor, through OpenVINO. We did not add a graphics card, a Coral stick, or a cloud GPU. Which detector you actually need.
The same camera streams the maker’s app already showsFrigate uses that address. You do not need a second special camera.
Detection set to catA hit is a visit. We cut the visit. We do not keep a 24-hour tape as the show.
The path she walked through the frameA tall phone crop can follow her instead of pinning the middle of an empty room.
Same computer cuts the ShortPython and ffmpeg. Wide 2560×1440 at 20 fps in. 1080×1920 Short out. Then YouTube Studio uploads it.

What a Frigate camera is for

A Frigate camera is not a leftover box on a shelf. It is a job in the room. Water watches the drink. Food watches the meal. Treat watches the quilt when a treat fires. Court watches the whole floor. Face is close enough to be a portrait. If you cannot say the job in one word, the camera is not ready to add.

Name the camera after that job. Do not keep a factory label, a room number, or the name of the last house it sat in. When a visit lands, you need to know which station she used without decoding leftover hardware.

  1. Install Frigate on a regular computer. No graphics card required. Start at frigate.video. Pick the install they recommend for a computer that can stay on. Ours is a Windows PC, and the Windows install has its own page. The Intel graphics inside the processor finds the cat; you do not need a graphics card. Finish their first-run until you can open Frigate in a browser on that same computer.
  2. Get each camera onto the home network first Plug the camera in. Add it in the camera maker’s own app until you can see a live picture there. Write down the stream address the app or the booklet shows — that is what Frigate will watch. Wired is steadier. Wireless is fine if the sitting room is close to the house box. The stream address, the sub-stream, and why we put go2rtc in the middle: camera setup. Which cameras speak that language at all: which camera.
  3. Add one camera per job In Frigate, add a camera, paste that stream address, and name it by the job: Water, Food, Treat, Court, Face. Point the real lens at that job before you save. If two cameras share a name, you will cut the wrong visit later.
  4. Turn on detection for cat Ask Frigate to look for a cat. Leave person and car off unless you have a reason they belong on the show. Only add the rooms you mean to publish. The exact objects: block, and why we do track person on the cat cams, is on the annotated config.
  5. Save visits, not a 24-hour tape Tell Frigate to keep a clip when a cat is found. You need the minutes she was in the job. You do not need the hours she slept off-camera. We keep three days of everything and three weeks of cat; the numbers are on the retention page.
  6. Treat a detection as a visit When Frigate says a cat was there, that is your cue to cut. Open that visit. Do not sit through the whole day.
  7. Keep her in a tall phone picture Use the path she walked through the frame. A phone picture is tall. A wide camera is wide. If you freeze a crop in the center, her ears walk out. Follow the path so the tall window stays on her.
  8. Review the visit as pictures Click through the stills from that visit. If they are empty room, do not post. Wait for the next visit.
  9. Keep Frigate for you The public site is the court and the Shorts. Send visitors to the live room or a finished cut, not to the Frigate window. The four public stills on the live site come through a read-only proxy that never exposes Frigate itself.

Frigate NVR watches. A visit is a detection. Pictures decide if the visit is real. Then you cut. How the cameras sit. How the cut is judged. The pipeline that cuts without a person.

The pages above are the plain-language layer. The pasteable layer, with every file as it runs: the WSL2 and Docker files, the full config.yml, and the OpenVINO detector over ZMQ.