Someone sets up a new AI assistant, connects it to their calendar, their file system, and a project database, and within a few minutes it is scheduling meetings, pulling up the right document, and answering questions about records that live nowhere near the model itself. No engineer was hired to wire any of that together. Compare that to how this usually went even a year or two earlier: a team wanting a chatbot to read from a company database would write bespoke code just for that one model talking to that one database, then repeat the entire exercise from scratch for the next tool, and again for the next model. Every pairing was its own project. The difference between those two experiences is a quiet piece of plumbing called MCP, the Model Context Protocol, and it is turning into one of those unglamorous standards that ends up mattering more than almost anything built on top of it.

The mechanism: one plug, many devices

Before MCP, connecting a model to an external tool or data source meant custom work for that exact combination. A model that could check a calendar in one application had no idea how to do it in another, because the integration had been hand-built for that specific pairing and did not transfer. Multiply that across every tool a company wanted to connect and every model it might want to use, and the number of one-off integrations needed grows fast, each one requiring its own code, its own maintenance, its own way of breaking when either side changed.

This is close to the situation computer peripherals were in before USB existed. A printer needed its own port, its own cable, and often its own driver, engineered for that one printer and that one computer. A scanner needed a different port and a different driver. A mouse needed yet another. None of it was interchangeable, and connecting a new device to a new machine was never guaranteed to just work. USB replaced all of that with one shared physical and logical standard: any compliant device could plug into any compliant computer, because both sides agreed in advance on the same interface. MCP is doing the same job for AI applications. It defines a common way for a model-based application to ask a tool what it can do and how to use it, so a calendar, a file system, or a search engine that speaks MCP works with any model-based application that also speaks MCP, and a new model swapped in on one side does not require rebuilding the connection on the other.

Why the plumbing question outranks the model question

It is easy to focus on which model writes better code or answers more accurately, because that comparison is visible and gets published in benchmark tables. But a model that cannot reach a calendar, a codebase, or a customer record is limited to whatever information happened to be in its training data or whatever a person manually pastes into the conversation. The practical usefulness of an AI application often depends less on the model’s raw ability and more on how many tools and data sources it can reliably reach without someone building a custom bridge for each one. A shared protocol turns that from an engineering bottleneck into a solved problem, the same way USB turned “will this device work with this computer” from a real question into an assumption nobody bothers to check anymore.

What the standard fight is really about

For more on the mechanism that lets a model request a tool call in the first place, one level down from the protocol itself, see Function Calling: How a Model Actually Uses a Tool. A later piece in this series looks at AI agents, which string these tool calls together into longer multi-step action. But the deeper pattern worth sitting with is this: the race to build the single best underlying model has settled into a crowded field of capable options from multiple vendors, with the gaps between them narrowing every few months. The more consequential contest now is over which connection protocol becomes the one everyone builds against, because whoever’s standard wins inherits outsized influence over the entire ecosystem of tools, applications, and habits that gets built on top of it. Plug formats have always outlasted and outweighed any single device that happened to use them, and the same is shaping up to be true here.