We’ve been thinking about AI development backwards. Instead of building “AI-powered” apps, we should be building tools that power AI. The Model Context Protocol lets you create small, focused tools that LLMs can use autonomously - just like Unix commands from 1978, but orchestrated by Claude instead of bash. I built a currency converter tool and watched an LLM decide when to use it, extract dates from user prompts, and combine multiple API calls without me writing any of that logic. It’s the same composable philosophy that made Unix great, applied to making AI actually useful in the real world.