Model Context Protocol
Every tool an agent can reach needs connecting code written by somebody. With no standard, that meant the same calendar integration written once per framework, and again when the next framework appeared. The Model Context Protocol is an open specification that replaces the multiplication with one interface on each side.
What the specification defines
Published by Anthropic in late 2024, it describes a host application, the clients it runs, and the servers they connect to, exchanging JSON-RPC messages over stdio or streamable HTTP. An initialisation handshake negotiates capabilities, after which a server may expose:
- tools — schema-typed functions the model can choose to call;
- resources — data addressed by URI and read into context;
- prompts — prepared templates a user can select.
Clients offer capabilities back, including sampling, so a server can ask the host's model for a completion instead of carrying a model of its own.
Why it matters
One server written for your database serves a desktop assistant, an IDE and your own agent script unchanged, and adoption well beyond Anthropic's own products makes that a working reality rather than a design goal. It also puts the trust boundary somewhere useful: the server holds the credentials and decides what it exposes, which is the one place scope can be narrowed for every agent at once. That same boundary carries the risk — tool output flows straight into the model's context, so a hostile resource is a prompt injection vector.
The Model Context Protocol, published by Anthropic in 2024, is an open standard for connecting agents to external systems: servers expose tools, resources and prompts over JSON-RPC, and any compliant client can use them. It replaces per-framework integration code with one interface on each side, and concentrates credentials and scope at the server boundary — which is also where injected tool output becomes a risk.
Railways in neighbouring countries were once built to different widths, so a train simply stopped at the border. The Model Context Protocol is an agreed width for AI assistants and the tools they use.
It is a published set of rules covering how an assistant asks for something to be done — look up a file, search a database, send a message — and how the other side describes what it has to offer. The rules are open, so anyone can follow them.
That means whoever looks after a company's systems can connect them once, and every assistant that follows the standard can work with them straight away, instead of each one needing its own hand-built connection.