Inspiration
We've all done it. Signed up for a $5 trial and got charged $20 the next month. Accepted terms you never read. Submitted the wrong file. The mistake was always right there on the screen, and we only noticed after clicking. We wanted something that catches it first.
That’s why we built Snoopsie: right before you click Pay, Send or Submit, it checks your entire screen and warns you in time.
What it does
Snoopsie lives in your Mac's notch. Press ⌥⌘K and it reads the page in front of you (a checkout, an email, terms, an assignment) and shows a small card of things worth a second look: a mistyped email, a yearly charge dressed up as monthly, a price that jumps after the first year, a box that's already ticked.
Each finding quotes the exact line, and "Show Me" takes you to it. Snoopsie never clicks anything for you. It points, and you decide.
How we built it
We built it in Swift and SwiftUI. Snoopsie reads the page through macOS Accessibility and hides card numbers and passwords before anything else looks at it. Then 91 quick built-in checks run, followed by Google's Gemma 4, running on your Mac, for the things rules miss. Gemini is optional, for very long fine print, and only ever sees the masked text.
Challenges we ran into
Our first version didn’t work as well as we hoped. It was too generic and didn’t really consider individual users, so everyone ended up getting the same broad advice. During early testing, we also ran into a lot of errors, like mistaking phone numbers for email addresses or flagging things that were completely fine. The small on-device model also struggled to catch unusual clauses hidden in long terms and conditions, and sometimes even made up issues that weren’t actually there. Another problem was that our indicator would sometimes cover buttons that users needed to click. We worked through these problems one by one. We made the checks more specific to each page and user, tested the system on real websites, and added a new test whenever we found a mistake. We also broke longer texts into smaller sections and gave the model more focused questions. To prevent false claims, we required every finding to include an exact quote from the page. Finally, we made sure the indicator checked its
Accomplishments that we're proud of
It works on real sites we never tuned it for, like the New York Times, Netflix, Canvas, and Discord. It has ~288 tests, and everything stays on your Mac by default: no account, no tracking.
What we learned
A warning app is only useful if you trust it, and only if it knows when to stay quiet. Fast rules plus a careful AI beat either one alone.
What's next for Snoopsie
A signed release, more language, and checking an AI agent's work before it clicks Pay for you.
