A Critical Look at MCP

Summary

A Critical Look at MCP "MCP is an open protocol that standardizes how applications provide context to LLMs. Think of MCP like a USB-C port for AI applications. Just as USB-C provides a standardized way to connect your devices to various peripherals and accessories, MCP provides a standardized way to connect AI models to different data sources and tools." ― Anthropic TL;DR I would like for this to turn out to be a skill issue on my part, and hope that I'm missing something. During the past month,MCP (Model Context Protocol), which would enable LLMs to become agents and interact with the world, has really been blowing up. The idea is straightforward: let's standardize an API for LLM/Agents to interact with the world and how to inform the LLM/Agent about it. Things are moving really fast, and IBM recently released their own "orthogonal standard" to MCP called Agent Communication Protocol (ACP), followed closely by Google announcing Agent2Agent (A2A). MCP Servers and Clients are being built and published daily, and can be found at sites like mcp.so and pulsemcp.com. However, I'm astonished by the apparent lack of mature engineering practices. All the major players spend billions of dollars on training and tuning their models, only to turn around and, from what I can tell, have an interns write the documentation, providing subpar SDKs and very little in terms of implementation guidance. This trend seems to have continued with MCP, resulting in some very strange design decisions, poor documentation, and an even worse specification of the actual protocols. My conclusion is that the whole suggested setup for HTTP transport (SSE+HTTP and Streamable HTTP) should be thrown out and replaced with something that mimics stdio... Websockets. Background About three weeks ago, I decided to jump on the MCP bandwagon to give it a try and see how it could be used in our own environment. I'm very much a person who wants to understand how things actually work under the hood before I start...