How MCP Servers Work

MCP Server Architecture

An MCP server is a lightweight process that exposes tools to AI agents through a standardized protocol. It acts as a bridge between an AI model and an external system — a database, an API, a file system, or any other service.

The Request Flow

When an AI agent wants to use a tool, the following sequence occurs:

1. Discovery — The agent queries the MCP server for its available capabilities
2. Selection — The agent (or the model) chooses which capability to invoke
3. Execution — The agent sends a structured request with parameters
4. Response — The MCP server executes the action and returns the result

This is similar to how REST APIs work, but specifically designed for AI agent workflows. The protocol includes metadata about permissions, parameter types, and expected outputs that help AI models understand how to use each tool correctly.

Capabilities and Permissions

Every MCP server declares its capabilities upfront. A capability includes:

• Name — A machine-readable identifier like query_accounts
• Description — A human-readable explanation of what it does
• Parameters — The inputs the capability expects
• Permissions — What access level is required

Permissions are critical for enterprise deployments. Each permission has a risk level (low, medium, high, critical) that helps organizations make informed decisions about granting access.

Running an MCP Server

Most MCP servers run as local processes alongside the AI agent. The typical setup:

npx @mcp/your-server

The server starts, registers its capabilities, and waits for requests. When running with Claude Desktop, you configure the server in your claude_desktop_config.json file, and it launches automatically.

Building Your Own

Building an MCP server involves implementing the protocol handlers for your service. The community has built servers for dozens of services — from PostgreSQL and Slack to AWS and Kubernetes. If your service isn't covered yet, you can submit one to the VaultPlane registry.