A web search and page reading tool for AI agents, shared with friends. It searches the web, reads pages, keeps the parts that answer your question, and can summarize them. No account needed.
Add an MCP server over streamable HTTP:
https://search.nkp.pet/mcpTools: web_search, web_contents, web_answer and a few single-page shortcuts.
Each address can make 20 requests a minute. All tools and HTTP endpoints.
Search results come from a self-hosted SearXNG. Pages are read in a real browser. Deciding which results and passages are relevant is done by Cloudflare-hosted Clef-Flash; it will move to a self-hosted Jev-like model at some point. Summaries and answers are written by Bonsai 2 27B, a 1-bit model run on a home server, not fine-tuned yet. Expect it to be slower and rougher than a big hosted model.
By default nothing is kept: your queries, the pages read and the answers are not stored.
If you want to help, you can opt in. Opted-in calls are saved (the query, the page excerpts and the
model's answers) and used for one thing: fine-tuning the model that writes the summaries, so they get better.
Send the header X-Contribute: 1, or use https://search.nkp.pet/mcp/contribute instead.
This is run by one person. The data is never sold or shared, and nobody reads through it; it only goes into training.
Please don't send anything private through it either way: it fetches pages and talks to model providers on your behalf.