What is an MCP Server?

An MCP server is a program that exposes a set of tools an AI agent can call — functions like search_issues or create_ticket — through the Model Context Protocol's standard interface. It's the "tool provider" side of MCP: the agent is the client, the server offers the capabilities.

The client/server split

The Model Context Protocol defines two roles. The client is the agent that wants to use tools. The server is the thing that offers them. A GitHub MCP server exposes tools to search code and open issues; a Stripe MCP server exposes tools to look up customers and issue refunds. Because the protocol is standard, any MCP-compatible agent can talk to any MCP server — you wire a capability once and it works across platforms. For the broader picture, see what is MCP.

What a server exposes

An MCP server publishes a list of tools, each with a name, a description, and a typed input schema so the agent knows how to call it correctly. Tools fall into two kinds. Read tools return data — searching, looking up, listing — and change nothing. Write tools mutate something — creating, updating, deleting. The distinction matters for safety: writes should require approval, reads don't. A well-designed server marks its mutating tools clearly so the agent platform can gate them.

Hosted vs custom servers

You can use existing servers or build your own. Many services offer hosted MCP servers, and platforms like AIML.chat ship a marketplace of ready connectors — GitHub, Stripe, Slack, Notion, and more — that install in a few clicks. When no connector exists for your system, you build a custom server that wraps your API, or use a custom-webhook option for simpler cases. See building a custom MCP server in 30 minutes for a walkthrough.

How an agent uses a server

When you attach a server's tools to an agent, the agent can call them mid-conversation. A support agent connected to a help-desk server can open a ticket; a technical agent connected to a GitHub server can file an issue. The agent decides when a tool is relevant, calls it, and uses the result in its answer. Reads happen automatically; writes pause for human write-confirm, so a person approves before anything changes.

Security considerations

Connecting a server means giving an agent reach into a system, so scope matters. Authorize each server with least privilege — a single repo, project, or read-only role where possible. Store credentials encrypted, never in logs. And keep an audit trail of tool calls so you can review what the agent did. These controls, plus write-confirm, make connecting a server safe. Read MCP security: SSRF protection for more.

Why standardization matters here

The quiet power of an MCP server is that it's reusable. Wrap your internal pricing API in an MCP server once, and any MCP-compatible agent can use it — today's platform and tomorrow's. That's a sharp contrast to the old way, where each AI integration was a bespoke project bound to one vendor. A server is an investment that doesn't evaporate when you switch tools, because the protocol, not the platform, defines the contract. For teams building real agent capabilities, that portability is the difference between a connector you maintain forever and one you write once.

Getting started

Connect your first MCP server on the AIML.chat free plan: install a connector, scope it, and attach its tools to an agent. Compare tiers on the pricing page, and read 10 MCP servers that make AI agents useful for ideas.