MCP server field guide

What Is an MCP Server? Model Context Protocol Explained

A practical explanation of MCP servers, hosts, clients, tools, resources, prompts, and the boundary between an AI application and the systems it can use.

9 min readEducational field noteReviewed 2026-08-26
For
Builders, technical leaders, and AI users who need a clear mental model before choosing or creating an MCP server.
Problem
MCP is often described as a connector without explaining who owns the connection, which capabilities are exposed, or what happens when an AI application calls a tool.
Useful outcome
Leave with a useful map of the MCP ecosystem and enough vocabulary to evaluate a server, client, integration, or build tutorial without treating them as interchangeable.

The route

Understand the actors before you connect a tool.

MCP becomes easier to reason about when the host, client, server, transport, and exposed capability each have a clear job.

Step 01
Define
Name the protocol, actor, and desired outcome
Step 02
Compare
Separate primitives, transports, and client behavior
Step 03
Choose
Select the smallest useful MCP route
Step 04
Verify
Check sources, permissions, and failure paths
MCP becomes easier to reason about when the host, client, server, transport, and exposed capability each have a clear job.

Workflow context: AI hosts / MCP clients / MCP servers / Tools / Resources / Prompts

MCP is a shared protocol boundary.

Model Context Protocol is an open protocol for connecting AI applications to external systems. An MCP server exposes a focused capability such as searching a knowledge base, reading files, calling an API, or taking a bounded action. The server does not become the model and it does not automatically receive authority over every system around it.

The useful boundary is explicit: an AI host chooses a server, an MCP client inside that host manages the connection, and the server implements the tools, resources, or prompts that the client can discover. The protocol gives those parts a common language; it does not remove the need for permissions, approvals, or application-specific policy.

  • Host: the AI application or environment the user operates.
  • Client: the connection component that speaks MCP to one server.
  • Server: the service or process that exposes capabilities.

Tools, resources, and prompts are different capabilities.

Tools are actions a model may request, such as querying an API or creating a record. Resources provide contextual data that an application can read or attach to a conversation. Prompts are reusable instructions that help a user or application start a known interaction. A good server exposes the smallest capability that makes the workflow useful.

That distinction matters for control. A read-only resource, an approved tool call, and a user-selected prompt do not carry the same operational risk. Treating every exposed item as an unrestricted action is a design mistake, especially when the server can reach customer data or production systems.

MCP is not an automatic trust layer.

Adding an MCP server does not make an external source accurate, safe, or authorized. The client still needs to decide which servers are trusted, which tools require approval, and which credentials are available. The server needs input validation, least-privilege access, useful errors, and a way to explain or record what happened.

It is also not the same thing as WebMCP. A remote MCP server is a service or process connected through an MCP transport. WebMCP is an emerging browser-native direction for pages to expose tools to web agents. The two can complement each other, but they solve different connection problems.

Use the model to choose the next article or build step.

If you need vocabulary, start with the architecture and primitives. If you need to build, move to a TypeScript or FastMCP server and then inspect it with MCP Inspector. If you need to connect a product, choose the client guide that matches your host and review its authentication and approval behavior before sharing credentials.

The reliable starting point is a narrow route with a visible outcome, a small tool set, and a named recovery owner. MCP gives the route a standard interface; the quality still comes from the design around it.

Reference material

Start with the platform documentation.

This field note is an educational guide. Platform behavior, availability, permissions, and plan limits should always be checked against the current vendor documentation.