What is a Model Context Protocol?

Prior to this standard, connecting a language model application to an external system, such as a database, a file store, or a SaaS API, meant writing a custom integration for that specific pairing, with no reuse across products. The protocol standardizes the shape of that connection instead: a client-server model where servers declare a set of tools for actions, resources for data, and prompts for templates in a common schema, and any compliant client can discover and use them the moment a connection is established, with no bespoke glue code required. Under the hood, messages are exchanged as JSON-RPC 2.0 over one of two transports — local subprocess communication for same-machine integrations, or an HTTP-based streaming transport for networked ones — with capability negotiation happening at connection time so client and server agree on what features both sides actually support before any calls are made. The practical effect has been to turn AI-tool integration into a build-once, use-everywhere problem: a company builds one server for its product, and every compliant AI application, not just the one it originally targeted, can connect to it without further integration work on either side.