Definition
The Model Context Protocol (MCP) is an open standard, introduced by Anthropic in late 2024, that defines a common way for AI assistants to connect to external tools, data sources, and systems. Rather than every AI application building a bespoke integration for every service, MCP provides a single protocol so that any compatible AI client can talk to any compatible server — much as USB-C standardised physical connections between devices.
Hosts, clients, and servers
MCP uses a three-part architecture:
- Host — the AI application the user interacts with, such as a desktop chat app, an IDE, or an agent framework.
- Client — a connector living inside the host that manages a one-to-one connection with a single server.
- Server — a lightweight standalone program that exposes specific capabilities, such as access to a filesystem, a database, a SaaS API, or an internal tool.
A single host can spin up several clients, each connected to a different server, letting one AI assistant simultaneously reach a code repository, a ticketing system, and a documentation store.
The core primitives: resources, tools, and prompts
An MCP server shares its capabilities through three primitives:
- Resources — readable data, like file contents or query results, that the model can pull into its context.
- Tools — callable functions that let the model take actions or fetch live data, such as running a query or creating a record.
- Prompts — reusable, parameterised templates the server offers to guide the model through a particular workflow.
This clean separation lets a server expose exactly what it wants the model to read, do, and follow — with the host and user remaining in control of what is actually invoked.
How a request flows
When a user asks the assistant something that requires external data or actions, the host’s model can decide to call a tool or read a resource offered by a connected server. The client relays that request to the server over the protocol (typically as JSON-RPC messages), the server executes it against the real system, and the result flows back into the model’s context to inform its response.
Why it matters
Before MCP, connecting M AI applications to N tools required building roughly M×N custom integrations. MCP collapses that into a single shared protocol, so building one server makes a capability available to every MCP-compatible client, and vice versa. This interoperability — combined with broad adoption across AI products — is why MCP is rapidly becoming the default plumbing for agentic AI that needs to reach beyond its training data into real tools and systems.