Forget MCP, Write CLI Apps

Over the last month or so, I’ve been usingClaude Code for assisting with development tasks1. One of the things that has confused me the most about the LLM-assisted programming ecosystem is theModel Context Protocol (MCP) landscape. In theory, MCP is a way to provide the LLM with data from an external system through a structured protocol. However,it sucks. It turns out that it is more efficient to have the AI just write code to analyse and manipulate the data from the external system directly. Rather than piping the data from the external system into the LLM’s context window and letting the LLM muddle through it, the LLM-written code can just do the necessary data processing. Computers have been good at running instructions on large datasets for a very long time and AI is not going to be a substitute for that.