MCP Architecture Explained: Hosts, Clients, Servers and What They Exchange

What is the MCP architecture?
MCP architecture connects the AI application (the host) to a server, a program offering functions and data, through a client, the application’s component for communicating with that server.
The client and server exchange messages over a local or remote connection. The server can offer tools for actions, resources for data, and prompts for reusable instructions.
Model Context Protocol (MCP) connects an AI application to outside tools and data. The base explainer covers what MCP is and why it exists.
Suppose you ask the assistant in your editor which issues are open in your repository, the place where your project’s code is stored. Issues are tasks or reports recorded in an issue tracker, a service that manages them. The architecture describes the parts that carry that request and return the answer.
The host, client and server
The host is the AI application you interact with, such as your editor’s assistant. The model is the language model the AI application uses to interpret your request and form an answer.
The host manages context across clients and keeps the full conversation; a server receives only the context it needs. The AI application passes information between the model and client.
The client is a component inside the AI application that communicates with an MCP server. The server is a program that makes capabilities, or available functions and features, accessible to the client. The AI application creates a separate client for each server to keep connections separate, as the MCP architecture specification describes.
Follow an issue request and its reply
The parts talk by sending messages: a request names a method, an action such as listing tools, and gets a reply. These messages use JSON-RPC, a standard message format.
Your editor’s client can optionally send server/discover first to learn the server’s supported versions and capabilities. It then sends tools/list to get the available tools, and the AI application shows them to the model. A tool is a function the server offers to perform an action or retrieve information.
The model selects the tool that finds open issues. In this example, the AI application asks you to approve the call, then the client sends tools/call with the repository and open-issue filter as arguments, the values the tool needs.
The server calls the issue tracker and returns its result to the client; the AI application gives it to the model, which answers you.
How messages reach the server
A transport is the way messages travel between client and server. With stdio, the client starts the server as a local program, and the client and server exchange messages through the program’s standard input and output. Your editor could use this route for a tracker server running on your computer.
With Streamable HTTP, the client reaches a server over a network at an endpoint, the address where the server accepts requests. The same issue request can go to a remote tracker server this way; the model and the roles stay the same.
What the server offers
By design, tools are model-controlled, resources are application-driven, and prompts are user-controlled; the specification does not require any particular interface for this. A tool does something when called, such as searching open issues; a resource is data the AI application attaches for context, such as the text of an issue.
A prompt template is reusable text that structures an interaction with the model: the tracker server could offer a template for summarizing open issues, which you select in the AI application.
Two parts of the specification cover your approval: the page on tools says the AI application should show tool calls and let you decline them, and the security principles say it must obtain your explicit consent before it runs a tool. MCP itself cannot enforce either rule; the AI application has to.
If you only wanted to see the parts and how a request travels, you can stop here.
Methods and request metadata
Every server must implement server/discover, which reports its supported versions and features; clients can choose whether to call it. tools/list lists tools and tools/call runs one; resources/list lists resources and resources/read reads one; prompts/list lists templates and prompts/get retrieves one.
Each request carries the protocol version and relevant client capabilities, or supported features, in _meta, a section containing information about the request.
When the server needs more input
If the issue tool needs a missing repository name, the server returns an input-required result asking for it. The client and AI application obtain your answer, and the client sends the original request again with that answer. The server can then look up the issues and return them. This request for input is called elicitation.
Authorization for a remote server
MCP authorization is optional: if the remote tracker server requires it, an authorization server issues the client an access token, a credential, after you authorize the request. The client presents the token and the MCP server checks it, as the MCP authorization specification describes; it is based on a selected subset of the OAuth 2.1 draft.
What changed in the current revision
The current revision, 2026-07-28, works differently from older revisions; check which versions your client and server support. The change log describes these changes:
- Previously, the client opened a connection with an initial exchange and the server remembered it; requests now carry the information needed to process them.
- Previously, a separate channel let a remote server push messages; change notifications, such as a changed tool list, now use a
subscriptions/listenrequest that the client opens. - Previously, servers could ask the client for information using their own requests; they now return an input-required result for the client to answer.
- Sampling, roots and logging are being phased out; they remain in the specification for at least twelve months after this revision’s release before they can be removed.
Where the boundaries matter
When it breaks: Check that the AI application asks for approval and review a server before granting it access. A local server runs with your account’s access, so review what it can reach.
For implementation, use Max's guide to building an MCP server.
Cheat sheet
The article on one page: open it full size, save it or share it.

AI-assisted content, human-reviewed. Images AI-generated. Editorial Standards · Our Editors