The second wave of MCP: Building for LLMs, not developers (vercel.com)

The second wave of the MCP (Model-Calling Protocol) standard shifts focus from exposing low-level API operations to building tools centered on complete user workflows tailored for large language models (LLMs). Early MCP implementations often mirrored APIs directly, creating thin wrappers that assumed developer-like state management and sequencing. However, LLMs interact differently: they start each conversation without memory, forcing them to repeatedly rediscover API toolchains and handle orchestration themselves—leading to inefficiencies and inconsistencies. To address this, the new MCP approach encourages designing tools that encapsulate entire workflows or user intentions in a single callable unit. For example, instead of four separate API calls to deploy a project, a single `deploy_project` tool handles the entire sequence internally, managing state, error recovery, and dependencies. This reduces complexity for LLMs, allowing them to generate conversational, context-aware responses rather than raw status codes. Teams adopting workflow-based MCP tools report greater reliability and smoother interaction, as these tools align better with how LLMs process tasks—focusing on outcomes rather than granular API steps. This paradigm shift signifies a crucial advancement for AI tool integration, emphasizing designing MCP tools around real user goals to maximize efficiency and natural language understanding.