Why I'm replacing MCPs with CLIs for my AI setup

Hey guys,

In my quest of clean context windows, I have find a way, sharing with you some learning.

MCPs are a great idea with a hidden cost: context bloat.

Every MCP you connect loads its full tool schema into your agent's context window. Doesn't matter if you're using 2 tools or 20, the whole thing gets serialized. I checked my setup recently and MCPs alone were eating 12% of my available context before I even started working.

The alternative I've been using: CLIs + skill docs.

Instead of an MCP that exposes 15 tools to your agent at all times, you build a CLI that does the same thing. Then you write a small skill doc, just a markdown file explaining what the CLI does and when to use it. Your agent reads the doc when relevant, runs the CLI via bash, done.

Benefits:

- Context only used when actually needed

- Works with Claude Code, Codex CLI, or any agent that can shell out

- Easier to debug (it's just bash commands)

- Can publish to npm and share with others

I've been converting my MCPs one by one. Down to 3% overhead now.

Last one I build for my PAI (Personal AI Assistant) meetgeek-cli

I can only suggest you guys give this a try!

Hope this help!

Quentin 🚀

1 comment