If you're running Claude with a dozen MCP servers and watching your context window fill up with tool schemas on every turn, this proxy collapses them down to three tools: mcp_search, mcp_call, and mcp_schema. The LLM queries in natural language to find the right tool, then calls it on demand. The proxy sits between your client and upstream servers like Google Workspace or custom APIs, handling discovery with BM25 lexical search and lazy loading tools only when needed. Pure JavaScript, no native dependencies. Most helpful when you're juggling 50+ tools across multiple servers and token overhead is eating into your actual work. Works over stdio or HTTP, and the included systemd service keeps it running in the background.
claude mcp add --transport stdio io.github.steveweltman-4q-tokenz uvx 4q-tokenz