Show the model text, not pixels.

KOALINT began as a simple frustration: agents that browse by screenshot burn tens of thousands of vision tokens to understand a single page — and then do it again on the next click.

The web is already structured. Links, fields, and buttons are right there in the DOM. So instead of a picture, KOALINT hands the model a compact, numbered map of exactly what it can act on — content-addressed, so the references stay stable as the page moves. Same understanding, a fraction of the cost, and it works behind your logins because everything runs locally on your Mac.

It's a real browser runtime — human and agent, headless when you want it, visible when you need it — with extraction, multi-agent decks, and human-in-the-loop safety built in from the start. No pixels sent, no data leaving the machine, no CAPTCHAs auto-solved.