Choose your flavour
Who are you hunting with today?
Someone stole my bike in London at the end of September 2026. The police did nothing, so I'm trying to get it back myself. Stolen bikes tend to get sold online, so a script on my Mac now checks every new bike listing on Gumtree and Facebook Marketplace and scores how much each one looks like mine.
Figures as of .
I can't scroll through thousands of bike listings a day, but a computer can. I built the checker with Claude's help. It runs on my laptop, and so does the AI that looks at the photos, so there's no cloud service and no bill per image.
A few times a day it opens Facebook Marketplace and Gumtree, newest first, and reads until it reaches listings it has already seen. It also searches the whole UK for words like "fixie" and "single speed", in case the bike has left London.
Each thumbnail goes through CLIP, an AI model that matches pictures to descriptions. It's asked a checklist of questions about my bike, one at a time, plus how similar the photo is to my own photos of it.
After each run, the top 20 new listings arrive on my phone as a grid of photos with their scores. If one scores 40 or more, the notification is marked urgent. A dashboard on my laptop shows every listing it's ever seen.
Colour · counts double
Is it blue and white, or black, red, silver, green…?
Type
A road or single-speed bike with thin tyres? Or a mountain bike, BMX, e-bike, kids' bike, or not a bike at all?
Saddle
White, black or brown?
Stripes
Coloured stripes on the frame, or plain paint?
Asking one question at a time works much better than giving the AI one long description, because it struggles to tell which colour belongs to which part of the frame. The answers are combined so that a bike has to match on everything to score highly. A blue mountain bike won't get far.
To set the bar, I scored my own photos of the bike against each other. They came out between 53 and 65. In early tests, the best real listing scored 44. So anything over 40 gets a look from me: a false alarm only costs a glance. It's a ranking, not a probability. So far nothing that scored high has turned out to be mine.
It's also deliberately polite. It waits between page loads like a person would, caps how much it reads each day and backs off for a day or more if a site pushes back. Gumtree blocked it on day one, so now it reads one Gumtree page every 20 minutes. For fun, there's also an over-the-top "scanning" animation version of the dashboard.
Have you seen this bike?
Newest first. If the bike turns up, it'll be at the top.
Just over five weeks in, the checker has looked at 156,029 listings. It finds around 4,400 new ones a day. 833 have scored 40 or more, which is enough for me to take a look. The highest score so far is 74.
The bike hasn't turned up yet. The checker keeps running.
The checker's first full run read 2,011 Facebook Marketplace listings in one go. Gumtree blocked it within a day, which is how it ended up with its very polite one-page-every-20-minutes routine.
