Model Context Protocol

Model Context Protocol is an open standard for connecting AI models to external tools and data sources through a common, agreed-upon interface. Rather than every AI application inventing its own custom way to reach a database, a file system, or a software service, Model Context Protocol defines a shared set of rules for how a model-based application and an external resource talk to each other. This lets any compatible AI application connect to any compatible tool or data source without bespoke integration work for each pairing.

The mechanics follow a client and server pattern. A resource, such as a document store, a business application, or a set of tools, is exposed through a server that speaks the protocol. An AI application acts as a client that can discover what a server offers, whether that is tools it can invoke, data it can read, or prompts it can use, and then interact with it through the standardized interface. Because the format of these exchanges is defined by the standard, the model can be given access to new capabilities simply by connecting to another compliant server, and tool builders can make their service available to many AI applications at once by implementing the server side a single time. This decouples the model from the specific systems it uses, replacing many one-off integrations with a common connector.

The term was coined by Anthropic in late 2024, when the company introduced and open-sourced the standard. "Protocol," from the Greek "protokollon," originally a leaf glued to the front of a manuscript and later meaning an agreed set of rules, is used here in its familiar technical sense of a defined way for systems to communicate. The name captures the purpose precisely: a protocol that standardizes how models are supplied with context and connected to the outside world. Being open, it was designed for broad adoption across different tools and providers rather than for a single product.

For a business, Model Context Protocol matters because it reduces the cost and fragility of connecting AI to the systems where real work happens. Integrations are often the hardest and most expensive part of deploying AI usefully, and a shared standard means a tool or data source built to the protocol can be reused across many AI applications instead of rebuilt each time. This encourages a growing ecosystem of ready-made connectors, so a business can plug an AI assistant into its knowledge base, its customer records, or its internal tools with far less custom engineering. It also lessens lock-in, since resources exposed through the standard are not tied to one specific application.

The nuances are worth understanding. A standard is only as valuable as its adoption, and its benefits grow as more tools and applications support it. Connecting a model to real data and actions raises the usual concerns around access control, security, and oversight, so the systems a model reaches through the protocol still need proper permissions and monitoring; a common interface simplifies connection but does not remove the responsibility to govern what the model can touch. The protocol is closely related to function calling, which is the underlying mechanism by which a model invokes a tool, and to grounding, since connecting models to authoritative sources helps keep their answers accurate. As a shared foundation, it makes tool-using and agentic AI substantially easier to build and maintain.