The second wave of MCP: Building for LLMs, not developers

The text discusses the challenges and solutions related to the implementation of MCP (Machine Communication Protocol) tools when interacting with large language models (LLMs). Initially, many teams created superficial MCP implementations by simply wrapping existing APIs to quickly claim support for the standard. However, this approach proved inefficient for LLMs, which lack the ability to recall past interactions or maintain state across conversations, leading to repeated and inconsistent orchestration of API calls. The text suggests that instead of exposing low-level API operations, tools should be designed around complete user intentions, such as deploying a project end-to-end, to streamline processes and reduce complexity for LLMs. This involves creating comprehensive workflow tools that manage the entire task internally and provide conversational updates, thus improving reliability and efficiency. The article emphasizes testing workflows manually before coding and adjusting tool design based on real user interactions to ensure complex workflows are successful on the first attempt, ultimately leading to better performance and user experience.