One week in July, India was politically charged, and so was I. I spent the entire week debating strangers and friends in DMs. I offended people I care about. I changed exactly nothing. The feed had decided what my week was about, and I had let it. So I built the sibling of FactLens: an Agent that judges my diet instead of the dish.
Nudger watches my own scrolling on X and Instagram. Every post that actually crosses my screen is classified by a local model into eight categories and written to a private ledger on my own disk. On the day of these screenshots, the ledger read: 180 posts sampled, 73 of them political. That is 41 percent of everything I looked at.
Nudger is a browser extension. It never posts, never blocks, never leaves the Mac mini it runs on. It watches only posts that genuinely enter the viewport, not the whole feed the platform served. A local language model classifies each post into one of eight categories: political, AI and tech, finance, entertainment, sports, other news, personal, other, with no cloud and no key. Every classification lands in a plain file on my own disk, one day per file, and minutes per platform are counted only while I am actually interacting. I set the threshold: which category, what percentage, after how many posts. When it crosses, a quiet full-screen note states the numbers and offers two buttons, back to it or snooze. Dismissing is always allowed. The point is the noticing.
This measures attention, not belief. The classifier mislabels edge cases, satire confuses it, and it only sees what crosses my screen, which is exactly the point but also the boundary. It cannot tell whether the political posts were worth reading. It can only tell me that I gave them 41 percent of my day, and let me decide whether that was a choice.
See FactLens, the sibling ยท Get in touch