What is an MCP server?
Without one, your AI can talk about your software all day and still can't press a single button in it.
Want to see it first? Book a demo
MCP server — An MCP server is a program that exposes tools, resources and prompts over the Model Context Protocol, so an AI application such as Claude, ChatGPT or Cursor can discover them and call them. Its purpose is to let any MCP-speaking AI use a system through one standard connection, with no custom integration per model.
Where the server sits in MCP
The Model Context Protocol splits the work three ways. The host is the AI app you use, such as Claude Desktop. Inside it, a client holds one connection to one server. The server is the program on the other end that owns the data and does the work.
A server offers three kinds of things. Tools are actions the model can call, like "create a deal". Resources are data it can read. Prompts are templates a user can pick. For a business system, tools matter most, because they decide what an agent can actually change.
You can run a server two ways. A local server starts on your machine as a subprocess the client launches, and messages travel over stdio. A remote MCP server runs as a web service: each message is an HTTP POST to a single MCP endpoint, so you can reach it from any machine you sign in on.
Host, client and server in the Model Context Protocol
| Role | What it is | Example |
|---|---|---|
| Host | The AI application you work in; it creates clients and handles your consent | Claude Desktop, Cursor |
| Client | One connection inside the host, paired with one server | The link from Claude to your CRM |
| Server | The program that exposes tools, resources and prompts | A CRM's MCP endpoint |
Roles and transports as defined in the MCP specification, version 2026-07-28, checked on modelcontextprotocol.io on 25 September 2026.
What happens when an agent uses an MCP server
- 1
The client sends a request
It carries its protocol version and capabilities.
- 2
It lists the tools
Each has a name, a description and its inputs.
- 3
The model picks one
It fills in the arguments from your request.
- 4
The server checks access
Is this user or token allowed this action?
- 5
It runs and replies
The result returns to the model to use next.
The mistake to watch for
What a CRM's MCP server looks like in use
SalesCrew runs one MCP server per client instance, over HTTP, with 140+ tools. Each tool has a twin in the app with the identical effect, and tokens are scoped by channel, profile and area like the app's own users. You can browse every tool on the MCP tools list.
Questions
- What is the difference between an MCP server and an MCP client?
- An MCP server exposes tools and runs them against real data. An MCP client lives inside the AI app and holds the connection to one server. The app, called the host, can run many clients at once, one per server you connect.
- Can one MCP server work with more than one AI app?
- Yes, that's the point of the standard. A server built once works with Claude, ChatGPT, Cursor or any other app that speaks MCP. You don't write a separate integration for each model.
- How does an MCP server control who can do what?
- The server checks each call against the caller's identity, usually a user login or a scoped token, before it runs. The spec also has hosts ask for your consent before any tool runs. Good servers log each call with who made it.
- Is an MCP server the same as an API?
- No. An API is a set of endpoints a developer codes against. An MCP server describes its tools in a form a model can read and pick from at run time, so an AI can use it without custom code. Many MCP servers call the product's own API underneath.