When does MCP make sense vs CLI?

Eric puts it well: LLMs already speak CLI fluently. MCP just adds transport, auth, and new failure modes on top of something that will work fine without it.

Claude Code is a CLI. That's the point. It's the same thing I see in so many Claude Code YouTube tutorials - the instructor running Claude Code inside VS Code — Electron wrapping Chromium wrapping Node.js just to reach a terminal. Three abstraction layers for zero benefit.

Same pattern at the protocol level.

Text in, text out, composable, inspectable. The Unix design approach is fifty years old and still relevant - it's the simplest and most direct method.