Code Is Better Than MCP

Current Model Context Protocol (MCP) lacks true composability and demands excessive context, making it less efficient than writing code.

Using CLI tools or direct code to automate tasks is faster, more reliable, and uses context more efficiently than MCP calls.

Automation at scale is best achieved with repeatable code pipelines, where LLMs generate code, execute it, and then validate results with minimal inference.

A proposed workflow involves an LLM generating transformation scripts, running diffs programmatically, and iterating through code rather than relying solely on inference.

MCP struggles with large-scale, inference‐heavy tasks, is hard to debug, and does not scale as well as code-based automation.

Future approaches should focus on combining code generation with targeted inference, better sandboxing, and APIs that enable fan-out/fan-in workflows.