The Model Context Protocol (MCP) is an open protocol that standardizes how applications provide context to Large Language Models (LLMs), enabling a standardized way to connect AI models to different data sources and tools. MCPs are compared to APIs, with the author suggesting they offer features like multiplexing and server push, similar to HTTP 2.0, and also include introspection/reflection capabilities, allowing for better tool discovery and usage. However, MCPs are not serverless and require a well-defined lifecycle, and their security relies on a holistic approach that includes running servers on remote machines with revocable bearer tokens. The author envisions MCPs as agents that can build things and adapt to users' needs, rather than mere appliances, and sees them as flipping the traditional "add LLM to something" paradigm, where instead of adding an LLM to a tool or application, you add the tool or application to the LLM.