AI Pulse · Under the Radar

A small open-source project called oc is making a practical claim for anyone building AI agents that need to read websites. Instead of passing a page’s raw HTML to a model, it renders the page as a numbered command-line view. The agent gets a cleaner, structured version of what a person would actually read.

The numbers posted in the project’s own benchmark are worth watching. Across six real web pages, oc used 1,936 tokens, compared with 177,685 for raw HTML and 16,402 for Jina Reader. It also says it was the only method to succeed on every test page. Those are project-reported results, not an independent benchmark, but the table and test details are available in the repository for anyone to inspect.

Why should an operator care? Web research agents can become surprisingly expensive when every page arrives as a mess of navigation, scripts, cookie banners, and hidden markup. If a tool can reliably strip that noise before the model sees it, your agent can handle more research, monitoring, and competitor tracking for the same budget. It may also make agent outputs more dependable because the model is working from a simpler page representation.

This project is only four days old and has 118 stars, so it’s far too early to treat it as proven infrastructure. Still, it points to a useful design principle: manage what information reaches the model, rather than just paying for larger context windows. This is the kind of thing we build into an AI command centre, where cost, source quality, and agent reliability need to be visible together.