A small MCP server and Node library that searches the web in the Chrome you already have, reads the best results and returns clean Markdown inside a hard token budget. There is no bundled browser, no API key and no LLM inside.
What it does
Works with any MCP client over stdio or a shared HTTP server, or as a dependency in code.
Extracts the main content by page layout, keeping headings, code, tables and lists, and dropping menus, ads and cookie banners.
Packs sections into the token budget, keeping the ones most relevant to the query.
Powers the web lookup step in whatcmd.
How it is built
- Search results are read three independent ways (layout, accessibility tree and selectors) and merged by vote, so a redesign rarely breaks it.
- Rate budgets per search engine are shared across every process on the machine, with longer pauses after a CAPTCHA.
- Bot checks are reported as blocked and never bypassed.
- No private-network access, and pages are read without your cookies.
Honest limits
- Search engines may block automated browsers.
- No PDFs yet, and Windows is untested.