Why CLIs Beat MCP for AI Agents — And How to Build Your Own CLI Army.

Six words. Thats what Peter Steinberger — the guy behind OpenClaw, 190,000 GitHub stars, freshly recruited by Sam Altman  — posted on X last month. And my immediate reaction was to screenshot it and send it to three dev friends with “ TOLD YOU SO ” in all caps. I’ve been building on Ubuntu for years. Every tool I use daily is a CLI. Supabase CLI, Vercel CLI, Docker, git, n8n — my entire stack runs from a terminal. When MCP servers started trending last year, I tried a few. They worked. They also ate 40% of my context window, crashed randomly, and added a dependency for something I could already do with a one-liner and a pipe. So when the most prolific open-source dev of 2026 says CLIs are the real interface between AI agents and the world — and OpenAI agrees enough to hire him — maybe it’s time to pay attention. TL;DR: MCP servers bloat your context window, add fragile dependencies, and solve a problem that doesn’t exist if your tools are CLIs 😂 Peter Steinberger built ~10 custom CLIs