Central téléphonique interurbain, Paris, 1937 — Someone has to understand the request before any connection gets made.
MCP
n.—the protocol that let a model ask a question instead of guessing.
The Model Context Protocol, introduced by Anthropic in late 2024, solves a problem that sounds small and isn't: before it existed, every integration between a model and an external tool—a database, a file system, an API—had to be built custom, once per model, once per tool. MCP standardizes that connection: a model-agnostic way for an AI system to discover what tools are available, what they do, and how to call them, without custom glue code for every pairing.
The shift it enables is less about raw capability and more about reducing friction at the connection points—the same kind of standardization that USB, HTTP, or SQL provided in their respective domains, arriving exactly when ad hoc integration had become the actual bottleneck rather than model quality itself.
What makes it more than plumbing is the second-order effect: once tool access is standardized, orchestration logic doesn't need to live inside the model's context window anymore. It can be externalized—into the protocol, into the runtime, into the tools themselves. That's the connection to the harness conviction: MCP isn't intelligence, it's the wiring that lets intelligence reach further without having to carry the wiring around in its own context every time.