Every AI tool integration was being written twice. Once for the model, once for the system it talked to. Four assistants and ten data sources meant forty bespoke connectors, each with its own auth, its own shape, its own maintenance.
Four plus ten is fourteen. That is the entire argument for MCP, and it is the same arithmetic that produced ODBC for databases and LSP for editors. Each side implements the protocol once, and the grid becomes a line.
The protocol separates three roles. The host is the application the user sees. The client speaks the protocol. The server exposes capabilities — tools to call, resources to read, prompts to reuse. Clear separation keeps each side replaceable.
There is a constraint hiding in that word exposes. Making a capability reachable says nothing about whether the model should be allowed to use it. That authorization boundary is a separate decision, and it is where tomorrow's card lives.
Standards don't win by being elegant. They win by turning multiplication into addition.