The text discusses the possibility of generating Model Context Protocol (MCP) servers directly from existing API specs using tools like openapisearch and Tadata. However, it highlights several pitfalls, including LLMs struggling with large amounts of choices, navigating complex JSON payloads, and dealing with ambiguity in API calls. The main argument is that MCP design should focus on workflows and tasks, rather than just exposing raw building blocks of the API spec. A more hybrid approach is suggested, where codegen is used as a starting point, but then tools are pruned and rewritten to provide clear examples and guidance for LLMs. Ultimately, higher-level tools that abstract away common multi-step workflows should be created to leverage the strengths of LLMs.