Solves the context window explosion problem that happens when AI agents naively fetch web pages. Exposes three MCP tools: webgate_query for full search cycles (SearXNG, Brave, Tavily, Exa, or SerpAPI backends), webgate_fetch for single URL retrieval, and webgate_onboarding for usage guidance. Strips HTML down to readable text with lxml parsing, hard-caps every response to prevent token blowouts, and optionally routes results through a local LLM for summarization. The real win is parallel fetch with BM25 reranking: a 13-page search that would dump 1.3 million tokens of raw HTML gets compressed to under 2,000 tokens of structured Markdown with citations. Built for Claude Desktop, Zed, Cursor, and CLI agents where a single unfiltered page can wipe your entire conversation history.
claude mcp add --transport stdio annibale-x-mcp-webgate uvx mcp-webgate