MCP Is Burning Your Tokens Before You Ask a Single Question

MCP (Model Context Protocol) injects every connected tool's name, description, and parameter schema into the LLM's context window on every turn, burning thousands of tokens before a single question is asked. As an alternative, CLIs can serve as the communication channel between agents and remote servers — for well-known tools like kubectl, gh, or AWS CLI, the LLM already knows how to use them with zero context cost. For custom or internal CLIs, lightweight 'skill files' (a few lines describing the tool) can teach the LLM what it needs, at a fraction of the token cost of MCP definitions. The trade-off: MCP offers centralized maintenance and standardized discovery across agent platforms, while CLIs require per-machine installation and distributed maintenance. The verdict: use CLIs for well-known tools, lean MCP for custom internal tools with few well-designed endpoints, and CLI with skill files when context footprint matters most. Both can coexist since the same backend API can expose MCP, HTTP, and CLI interfaces simultaneously.