mcp-cli is a lightweight CLI tool that solves the context window bloat problem in AI agent workflows by enabling dynamic, just-in-time discovery of MCP (Model Context Protocol) tools instead of loading all tool definitions upfront. In a benchmark scenario with 6 MCP servers and 60 tools, it reduces token consumption from ~47,000 to ~400 tokens — a 99% reduction. The tool features a 3-subcommand architecture (info, grep, call), connection pooling with a lazy-spawn daemon, glob-based tool search, and tool filtering via allowedTools/disabledTools config. It integrates with AI coding agents like Claude Code and Gemini CLI via system prompt instructions or Agent Skills definitions, and compiles to a single binary built on Bun.
