Classify web pages before adding them to an AI agent's context

A service that classifies web pages from their URLs and returns only the label answers with probabilities so an agent can decide whether to open a page without spending context tokens. Send a POST to the root endpoint with a URL, a question or a set of labels and receive the best-fitting label and probabilities; a separate System One endpoint accepts multiple named yes/no questions about one page and returns probability scores for each. Practical uses include checking links before an agent follows them (catching login walls, cookie notices, 200-OK error pages), filtering search results in parallel to read only relevant links, verifying whether a linked page still documents a specific command, and pre-sorting user-submitted links for review. Example client patterns include simple cURL requests and concurrent classification with a thread pool.

Operational details and limits: the service reads only public pages, does not follow links, and uses the page parts most relevant to the question; addresses embedded in input are read (up to five pages per request). Pages can self-assert labels, accuracy on arbitrary pages is unmeasured, and users should validate labels on known examples. Billing is by units per page with a 1,000-units-per-IP-per-day free tier (covering on the order of a few hundred checks depending on labels); failed reads return a 502, the failed page’s units are kept and the rest refunded. The system is open-source (syn) and keys come from sifty.dev/account.