Peekoo

Turn any photo into a desktop pet that climbs your windows

Windows

Mac

Artificial Intelligence

Hi Product Hunt 👋 I've had Bongo Cat on my desktop for years. So when a Codex desktop pet started going around, I wanted one immediately — spent three hours putting mine together and ended up with a little character that sat there and did nothing. Looked right. Wasn't alive. So I built Peekoo. It turns a photo you already have into a desktop pet, in about a minute. The first thing I did with it was point it at my dog, who was wearing an Argentina jersey at the time. A minute later he was standing on my screen — jersey and all. He's been living down there ever since, and I did not expect to enjoy that as much as I do. That's him in the first image. Getting there took a lot longer than a minute, though. The thing I got wrong first: I assumed everyone wanted something cute. They don't. Roughly half the early testers uploaded someone they were annoyed at. So the generator takes an identity and an affection level, and those actually steer the art. Set it high and the model is told to make them softer, rounder, more endearing. Set it low, or tag them as your boss, and the instruction flips: draw them goofy, slightly pathetic, a bumbling little sidekick — funny, never mean. Same photo, two very different characters. There's a groveling pose in the default set for exactly this reason. It gets used a lot. 1. One character, 23 images, no memory between them. A full pet is 23 generated frames — idle, walking, typing, napping, climbing, parachuting. The model remembers nothing between calls, so every frame is a fresh chance for your cat to change breed or come back 30% larger. Identity was the easy half: a baseline portrait goes into every subsequent call. Size was the subtle killer. Bounding boxes are useless here — a curled-up nap and a full-body stretch aren't comparable that way. Normalizing on silhouette area instead (√area as the scale key) makes both read as the same animal. Boxes lie; mass doesn't. 2. Not everything has four legs. Early prompts quietly assumed a humanoid or a quadruped. Then someone uploads a spider. An octopus. A ghost. Every "rest its front paws on the keyboard" becomes a contradiction, and the model resolves contradictions by inventing limbs. Now the prompt names the intent and hands the anatomy back to the subject: substitute whatever limbs it actually has — tentacles, a tail, fins, just leaning forward — and match the reference's limb count exactly. Never bolt on human hands. 3. A flat image isn't a body. To make a sprite behave physically you need to know where it touches the world, and that point moves per pose. Standing: the soles. Clinging to a wall: the palm, on the right edge. Hanging from a ceiling: the knuckles, on top. So every pose carries its own contact anchor, found by scanning alpha rather than trusting the layout. Miss that and the pet teleports every time it changes state. It did, for weeks. Everything after that was just physics — reading real window rectangles, skipping edges hidden behind other windows, leaping to catch a ledge because there's nothing to hold below a window's bottom. Native on macOS and Windows, ~11 MB, clicks pass right through it. Free to start. Curious: whose photo would you turn into one, and would you set the affection high or low?