The Digital Press

Armin Ronacher critiques the current Model Context Protocol (MCP) for automating coding tasks, arguing that generating and running code is often more efficient and reliable than relying on inference-heavy MCP approaches. He demonstrates how using LLMs to generate code for repetitive tasks allows for better validation, scalability, and trustworthiness compared to direct tool invocation via MCP.

Why it matters: Code generation enables scalable, repeatable automation with better reliability and verifiability than inference-based MCP tool usage.

The big picture: MCP relies heavily on inference, which limits composability and context efficiency, making code-based automation preferable for many tasks.

Quick takeaway: Using LLMs to write and judge code—rather than direct tool calls—can reduce errors, speed execution, and scale automation effectively.

Commenters say: Many agree with prioritizing code and shell-based tools over MCP, emphasizing reliability, token efficiency, and the importance of deterministic interfaces.