Jevcrawl wants your AI agent to stop paying an LLM to pick every link
The new MCP crawler uses hosted Jev scores to select links and verbatim passages, with an author-reported cost advantage on a small web-navigation benchmark.
A smaller model chooses where to browse
Jevcrawl has released an MCP server and command-line crawler built around a specific idea: use a small decision model to pick relevant links and passages, then give the answering LLM only the retained source text. Version 0.1.0 arrived October 13, followed by a Show HN submission later that day.
Its “no LLM in the crawl loop” framing does not mean model-free or fully offline browsing. The default setup calls TypeSafe’s hosted Jev model to score page content and links. Returned passages are copied verbatim and include URLs, heading paths and character offsets, so an agent can inspect where the text came from.
The author’s cost claim comes with a defined test
In an October 12 run of 50 WebWalkerQA tasks, the author reports 40% correct answers at $0.025 per task, against 30% and $0.194 when Claude selected links. The README says the systems shared a 30-page budget, answering model and judge. These are the project’s own small-sample results, not independent replication or a promise of savings on every website. A separate 20-task Crawl4AI comparison measures page retrieval and should not be read as the same accuracy experiment.
For builders, the release offers bounded crawl, scrape and map operations under Apache-2.0. Page text, link labels and URLs are sent to TypeSafe for scoring. Jevcrawl does not run JavaScript, log in or bypass paywalls and CAPTCHAs, and it does not read robots.txt or impose a rate limit beyond four concurrent page fetches. Its scoring design also does not establish blanket prompt-injection immunity: the documentation acknowledges that injected text can bias scores.