What is the Model Context Protocol (MCP)?

How does MCP work?

An AI application, called the host, runs an MCP client that talks to one or more MCP servers. Each server describes what it offers, the client discovers that and presents it to the model, and the model can then ask for a tool to be used.

What can an MCP server offer?

It can offer tools, which are actions the model may request, such as searching a ticket system. It can offer resources, which are data the application can read, such as a file or a record. It can also offer prompts, which are reusable templates a user can choose.

Why does it matter?

It turns an N-by-M integration problem into N plus M, because each application and each tool implements the protocol once. It also puts tool access in one place where it can be reviewed, instead of scattered through application code.

Is it safe to connect any MCP server?

No. A server is software that runs with real access, so treat a third-party server as untrusted code until it has been reviewed. Give it the least privilege it needs, keep credentials out of prompts, and require human confirmation for actions that change things. Content a server returns can also carry hidden instructions; see what is prompt injection.

How do you get started?

Run a well-known server locally against a test dataset, and read the tool descriptions the model will actually see. Limit its permissions, widen them only when needed, and log every tool call so you can see what the model did.

IICTP is an independent technical credentialing body.