Librechat requires scraper and reranker configured in addition to SearXNG before the "web" tool is presented to the LLM. I understand the reasoning behind it but I think it should be the administrator's choice if they want to only use SearXNG.
I went around this limitation by exposing SearXNG via MCP. One problem I'm having is that the context window fills very quickly when an agentic model like GLM-5 uses the search and URL read tools. I'm not sure if this problem happens with the "official" web tool in Librechat.