Connects to Claude Desktop, Cursor, or any MCP client and exposes three tools: fetch_clean for single URL retrieval, fetch_clean_batch for multiple URLs, and refine_prompt for optional query cleanup. Uses Python libraries to strip HTML cruft before pages hit your context window. The numbers are wild: a 704K token Yahoo Finance page drops to 2.6K tokens. Runs locally with no API keys or models. You'd reach for this when your agent is burning tokens on navigation bars and ad scripts instead of actual content. Works standalone as an HTTP proxy or drops into LangChain as a custom tool. Ships with caching so repeat fetches are instant.
claude mcp add --transport stdio xelektron-token-enhancer uvx token-enhancer