How is MCP built?
MCP connects an AI app to servers through one client per server. By design, servers get only the context the host app shares with them.
MCP architecture is how MCP is built. A host app creates one client for each server, and each client talks to one server. Servers supply tools, resources and prompts, and are designed not to see the whole conversation.
Also called MCP server architecture, MCP architecture overview, how MCP works, MCP host client server.
MCP has two layers. The data layer defines the messages between clients and servers, written in JSON-RPC 2.0. It covers capabilities, tools, resources, prompts and notifications.
MoreLess
The transport layer carries those messages. There are two standard ways. Stdio is for local servers on your machine. Streamable HTTP is for remote servers and supports sign-in with tokens, with OAuth recommended.
The docs use Visual Studio Code as the example host. It makes one client for the Sentry server and another for a local filesystem server. The Sentry server is remote, and the filesystem server is local.
Related:MCP hostMCP clientMCP serverLocal MCP serverRemote MCP serverModel Context Protocol (MCP)
MCP architecture has three roles. The host is the AI app, such as Claude Desktop or Claude Code. The host creates one client for each server it connects to, and the server is the program that provides tools and data. The one thing to know first: the host keeps the full conversation, and the spec says servers receive only necessary contextual information.
The spec's design principles say servers should not be able to read the whole conversation or see into other servers. The host enforces those boundaries. It also handles your approvals and sign-in decisions.
Servers can offer three things: tools, resources and prompts. They can also ask you for input through the app, for example with elicitation. The latest spec, 2026-07-28, describes MCP as stateless: every request carries its own protocol version and capabilities.
Good for
- Anyone asking what is happening when they add a server
- Admins judging what a server can and cannot see
- Builders planning a server or an app
Comparison
The host coordinates the clients, and each client connects to one server.
| Option | What it is | Example | How many |
|---|---|---|---|
| HostThis page | The AI app | Claude Code | One or more clients |
| Client | Part of the host, one per server | Inside the app | One per server |
| Server | Supplies tools and data | Sentry, filesystem | One or more per host |
Questions
How does MCP work?
An AI app (the host) creates one client for each MCP server you connect. The client asks the server what it offers, such as tools, resources and prompts, and passes the results to the app.
What are the parts of MCP architecture?
The host, the client and the server. The host is the AI app. The client is a piece inside it that handles one server. The server supplies tools and data. See MCP client for more.
Is an MCP server local or remote?
Can an MCP server see my whole conversation?
Not by design. The spec says servers should not read the whole conversation or see into other servers. They get only the context the host passes them.
What messages does MCP use?
JSON-RPC 2.0, the same format on every transport. You do not need to know it to use MCP, since the app handles it.
Is MCP stateless?
The latest spec, 2026-07-28, says so. Each request carries its own protocol version and capabilities, so it does not rely on an earlier one.
What can a server offer?
Three kinds of things: tools that act, resources that give context and prompts that are templates. Details are on the pages for each.
Does MCP decide how the AI uses the data?
No. The docs say MCP only covers how context is exchanged. How an app uses a model or handles the context is up to the app.
Sources
Docs, issue trackers and discussions this page is built from.
Show all 2Hide
- Architecture overview · modelcontextprotocol.io
- Architecture · modelcontextprotocol.io