How do I make YouTube Shorts from a live cat cam?
Dress one small room so a cat already has places to drink, eat, take a treat, cross the floor, and sit at a bowl. Point a camera at each job. Let a watcher find a real visit — do not sit in a chair and call “action.” Cut that visit in the order it happened. Open a beat after she arrives so frame one is already a cat. Close as she leaves. Dissolve the last frames into the first so the loop does not flash empty furniture. Write one short line at the bottom of the frame. Post it as a YouTube Short. The same loop works if your public site is a walkable 3D court: the court is how people watch live; the Short is how they find you.
What is Frigate NVR and how do you use it for a cat?
Frigate NVR is open-source camera software people use at home. You install it from the Frigate project. Add one camera per job in the room — water, food, treat, court, face — and name them by the job, not by leftover hardware names. Turn on detection for cat, and only in the rooms you mean to publish. Treat a detection as a visit. That visit is your cue to cut, not a day-long tape. Use the path she walked through the frame to keep her in a tall phone picture; a frozen center crop throws most of a wide camera away. Review the visit as pictures. If the pictures are empty room, do not post. Never put the Frigate screen on a public page.
How do I set up a live cat camera?
Pick one room. Do not cover the house. Mount cameras at cat height and give each one a reason to exist: a high water shot, a food wide that keeps body and bowl, an eye-level treat shot, an overhead of the whole room, and a face at the bowl for the opener. A cat at a bowl has her head down — aim slightly high of center so the face stays in frame. Publish a website people can open and look around. Refresh stills or streams from those cameras so the room feels alive. Do not model bedrooms or people. The working example is the live court, not a wall of security tiles.
How do I use several cameras in one cut?
Stay inside one visit. Keep every shot in the order it happened. We once shipped a cut that ran fourteen-point-nine seconds backwards because shots were sorted by the wrong clock. Now we refuse any cut where a later shot starts before the earlier one ended. Start about half a second into the visit so frame one already has a cat — not an empty establishing shot. End when she leaves. Do not hold on furniture. Loop by dissolving the last frames back to the first. Wide cameras on a phone become a tall slice. If you crop only the middle you lose the cat; pull back just enough to keep ears and set.
What is a cat treat dispenser doing on a live site?
We bought an ordinary dispenser that already had a remote. Then we bought a 5-volt dry WiFi switch — not a 120-volt one — and soldered it across the treat button inside that remote. Five volts. Dry means the switch only joins two wires — the same job as a fingertip. Aim Treat Cam at the quilt first. Sell the treat on your own shop. A payment pulses the switch. The dispenser hears the click it always heard. Do not force the cat. If she ignores it, that is still the joke. This making-of page does not take money.
Do you need a GPU to run Frigate for a cat cam?
No. We run Frigate on a regular Windows PC with no graphics card. The Intel graphics built into the processor finds the cat, through OpenVINO, at about 12 milliseconds a look across seven cameras; before August 2026 the processor itself did it, and that worked too. That same computer cuts the Short with Python and ffmpeg, then uploads through the YouTube API. You do not need a gaming GPU, a Coral stick, or a cloud farm. If the PC can stay on and see the cameras, it can do this. Which detector you actually need, and how the built-in graphics does it under WSL2.
Do you need special studio machines to copy Helen?
No. We used normal machines, no graphics card, and made them do amazing things. Seven ordinary home-network cameras, five jobs, send wide 2560 by 1440 pictures at twenty frames a second. A regular Windows PC that stays on runs Frigate, with the processor's built-in Intel graphics doing the detection — no graphics card, no Coral stick, no cloud farm. That same computer cuts the Short with Python and ffmpeg, then uploads through YouTube Studio. A white PetSafe dispenser already had a remote. A 5-volt dry WiFi switch — not a 120-volt one — got soldered to that remote so a shop payment can press it. The court is a website people walk. Copy the jobs, not a truck full of gear.
Is Helen a dilute tortoiseshell?
Yes. Helen is a dilute tortoiseshell calico rescue from the Woman’s Club of Winter Park. Not a tabby. The muted coat is the dilute. The patches are the tortoiseshell. The white is why people also say calico. The ear notch is from the rescue, not from the set. We write her as a monarch because the gap between a decree and an ordinary cat is the joke. She is never made to act.
How does a 3D interactive cat court work?
Measure one sitting room. Give the cat five reasons to go somewhere: water, food, treat, a crossing, and an eating close-up. Dress the real room so a camera already has a job when she does the ordinary thing. On the website, rebuild that room as a place people can turn and look — not a grid of camera tiles. Put each viewpoint where the real camera sits so visitors understand the geography. Refresh a still or a stream into each viewpoint so the 3D room feels alive. Face Cam may be a curated frame if it is not a public stream. Do not model bedrooms or people. Halloween dressing is scenery. The cat is not a prop.
How do you post automatically?
Decide four daily doors and keep them. Near each hour, Frigate finds a real visit. A cut is built in time order from the cameras that saw her. Write one decree: ten words or fewer, bottom of the frame, nothing on her face. Skip the hour if she barely moved. Rewrite the title if that sentence already ran this week. In the last seconds, ask them to give her a treat, show the live domain, then loop. Upload as a YouTube Short. Category pets. Not made for kids. Do not use a greater-than sign in the description — YouTube rejects it. Then look at the posted frames on a phone. If the avatar or thumbnail still says an old brand, fix the picture. Ours had that problem.
Why four posts a day?
Four doors is enough to stay in the Shorts feed without inventing visits. Ours are five in the morning, seven in the morning, five in the evening, and seven in the evening, Eastern. Each hour has a usual station — water, food, the whole court, or her face — so you are not hunting a random clip. If she barely moved, that door stays empty. A still clip trains people to swipe. A flood of empty rooms does the same. Four honest visits beat twelve quiet ones.
Can I copy this without your computers?
Yes. Start with cameras at cat height, then Frigate from frigate.video, then the cut rules, then YouTube Studio. Rebuild your room as a place people can walk. The steps are on this site.
How do I write a title for a cat YouTube Short?
Write one decree: ten words or fewer. Put it in the bottom third. Leave the middle and the top of the frame to the cat. Nothing covers her face. A longer season chapter can keep the same voice and spend it across four timed lines. If that sentence already ran this week, write a new one before you post. A reason line and a wordmark fighting the decree is how you lose the first half-second.
How many cameras do I need for a cat court?
Five jobs in one room. Water from high, over a fountain or a glow. Food wide enough for body and bowl. Treat at eye level, close enough to read whiskers. Court as an overhead of the whole sitting room — not a leftover scratch pad. Face head-on at the bowl, because that is the opener. A single wide cannot do a drink, a meal, a treat, a walk, and a face. Mount them low. Do not cover the house.
What if the cat does not move?
Do not post a still clip. Leave that door empty and wait for the next one. Look at the pictures before you trust a success message. If the frames are empty furniture, the visit is not a Short. The same rule applies to a treat drop she walks past: you may have an event in the room, but you still need a cat in the cut before you publish.
Where can I watch the live cat camera and the Shorts?
The performing site is the live cat camera and the walkable court. The channel is where the Shorts land. This domain teaches the method. If you came here looking for the show, you want the live site. If you came here to copy the method, start with cameras, Frigate, the cut, and four doors a day.
The deeper pages, one question each
What is Frigate NVR, in one paragraph?
Frigate is open-source software that watches the cameras you already own, runs object detection on your own computer, and keeps clips of the moments something it was asked to find walked into the picture, instead of a 24-hour tape. It is not a cloud service, not a camera brand, and not a phone app. Three ideas cover it: a detect stream and a record stream, a detector, and retention.
Can you run Frigate NVR on Windows?
Yes, and we do, 24/7, on an ordinary desktop. Ubuntu inside WSL2 with systemd turned on, Docker Engine inside that distro, Frigate as one container, recordings on a normal NTFS drive, and the Intel graphics reached through WSL2's GPU passthrough. The things that break are the WSL VM idling out, Windows Update reboots, and a boot task that refuses to run as SYSTEM; each has a fix.
How do I add a camera to Frigate?
Create a camera account in the maker's app, give the camera a fixed address on your network, and write down its two RTSP paths: the full-resolution main stream and the small sub-stream. Frigate detects on the sub-stream at 640×360 and two frames a second, and records the main stream untouched. We put go2rtc in the middle so the camera is pulled once and everything else reads from that.
Can I see your whole Frigate config?
Yes. Eleven cameras, seven detecting cat and person, one spare that is switched off, three house cameras that only record. The file is on the site with the passwords and addresses replaced by placeholders and every section explained: the ZMQ detector, the model block, retention at three days of everything and twenty-one of cat, the cat filter thresholds, and a copy-this block for a new camera.
Do I need a Coral for Frigate?
Probably not. A Coral was the standard answer for years, but the Intel graphics built into most desktop processors runs Frigate's model through OpenVINO fast enough for a house full of cameras; ours does about 12 milliseconds a look. Buy a Coral if you already own one. Buy a real GPU only if you have dozens of cameras or want bigger models.
How do you run Frigate detection on an Intel iGPU under WSL2?
Frigate's container cannot see the iGPU under WSL2, but a plain Python process on the WSL host can. So the detector runs outside the container and Frigate talks to it over its zmq detector type: a 170-line OpenVINO server that answers each frame with the block of detections Frigate expects. It has answered more than three million frames without a crash.
How much disk does a 24/7 camera setup use?
Ours writes about 107 gibibytes a day across nine recording cameras (16.4 for each always-on 1440p camera), and Frigate counts about 420 gibibytes of recordings on a 3.7 terabyte drive. The trick is the retention shape: keep three days of everything and twenty-one days of the moments a cat was actually there. The page has worked examples for four cameras and a one-terabyte drive.
How do you put a Frigate camera on a public website without exposing Frigate?
The public only ever sees one fresh still per camera, fetched through a small read-only proxy that knows the names of the four public cameras and refuses everything else. The proxy reaches Frigate over a private network; Frigate's own screen, the video streams, and the house cameras cannot be reached from the site at all.
How do you stream a security camera to YouTube 24/7 without OBS?
With ffmpeg on the computer that already has the streams. Each camera is restreamed once, a compositor stitches four into a 1080p mosaic and encodes it on the processor's built-in graphics, and a second script sends one close camera by itself. Two rules matter: re-time to a constant frame rate with an audio track, and watch the network connection rather than the process, because YouTube can drop you while ffmpeg keeps running.
Which cameras work with Frigate?
Any camera that publishes an RTSP stream you can open in VLC. Tapo does, with a camera account; so do Reolink, Amcrest, Dahua and Hikvision. Arlo, Nest and Ring do not, because they are cloud-locked. We run Tapo C120 and C210 cameras on Wi-Fi in one room close to the router; if you have walls or more than ten cameras, run cable.
How does paying for a treat press a button in your house?
The shop creates a Stripe payment. When it succeeds, Stripe sends our server a signed webhook. The server checks the signature, makes sure this payment has not already dropped a treat, checks a daily cap and a cooldown, then posts once to Home Assistant on the computer at home over a private tunnel. An automation turns on one channel of a Sonoff relay, and that channel is the switch soldered across the remote's button.
What hosting do you use, and how do you deploy?
One small rented server with a single processor runs both websites behind Caddy. The show site is built into a fresh timestamped directory and goes live by moving one symlink, so a bad release is undone the same way. Git is the only source of truth; nothing is edited on the server. A curl after every deploy is what counts as success.
How is the 3D court built?
It is a three.js scene in an ordinary Vite site. Four of the walls are live camera frames refreshed one at a time so the picture never hitches, the cat is a generated mesh compressed to 738 kilobytes, and a resolution loop follows whatever machine is looking at it. The whole site is 11 megabytes, down from 92.
How do the Shorts get cut and posted automatically?
A script polls Frigate for cat events, groups events across cameras into visits, builds a shot list in time order, pulls each shot from the continuous recording, crops it to follow the cat, has a model describe the frame and write the decree, burns the caption on, and uploads. It refuses a backwards assembly, posts at most one per run, and never publishes from a camera it does not know.
Did AI make this, and which AI did you use?
AI models do specific jobs here: describing a frame and writing a ten-word caption for each Short, generating the 3D cat from a photograph, and a great deal of the code, written with ChatGPT, Cursor, Grok, Claude, Gemini and DeepSeek at different times. A person still decides what ships, reviews every caption, and fixes what the generators get wrong. The site says which tool did what, and what it did badly.
How do you pick which AI model to call, and what happens when one fails?
The pipeline never names a model. It asks for a role, vision or caption, and a small resolver keeps an ordered list of models across more than one vendor for each. If the first does not answer, the next is tried; if none answer, a hand-written line in the voice is used and the Short still goes out. The day one vendor's billing limit took the channel dark is why it works that way.