The Model Context Protocol (MCP) has sparked fierce debate in the AI development community. Defenders praise its ability to connect AI clients like Claude Desktop to external tools in minutes, while critics point to reliability issues in agent codebases. Both sides are right — they're just describing different use cases.

MCP solves a real problem by standardizing tool integration across AI clients. Without it, every AI tool would need custom integrations for every service — a costly N×M problem. The protocol works exceptionally well for non-developers wanting AI assistants to access calendars, update tickets, or query databases without writing code. It also excels at documentation access, letting agents pull specific pages on demand rather than bloating context windows with entire docs sites.

The controversy stems from using MCP for programmatic tool calling in agent loops. When agents invoke tools directly via MCP, they operate in their weakest mode — LLMs were trained on billions of lines of real code, not synthetic tool-call JSON syntax. Cloudflare's engineering team found that having agents write code to call MCP tools instead of invoking them directly is far more reliable and reduces token overhead by 98.7%. For most developers building single applications, MCP's overhead — running server processes, context window bloat, and protocol complexity — outweighs any benefits.