What Is an MCP Server? The Plain-English Definition

An MCP server exposes your tools and data to AI models like Claude, ChatGPT, and Cursor through a standardized protocol.

TL;DR

An MCP server is a program that exposes tools, resources, and prompts to AI models via the Model Context Protocol. Two transports: stdio (local) and HTTP/SSE (remote). Auth via bearer tokens / OAuth, tenant-scoped, with HMAC-signed webhooks.

What Is MCP Server?

A program that implements the Model Context Protocol to expose tools, resources, and prompts to MCP-aware AI clients. Runs locally over stdio or remotely over HTTP/SSE.

Key Takeaways

  • Server exposes; client consumes
  • Two transports: stdio (local subprocess) and HTTP/SSE (remote)
  • Tools have JSON-Schema input definitions for safe calling
  • Production servers must scope tokens, sign webhooks, and log every call

Direct answer: An MCP server is a small program that exposes a set of tools, resources, and prompts to AI models through the Model Context Protocol. It's the bridge between an AI client (Claude, ChatGPT, Cursor) and a real-world system (your CRM, your database, your file system). If MCP is the standard, the MCP server is the plug.

Definition

MCP Server: A program that implements the Model Context Protocol to expose tools (callable actions), resources (readable data), and prompts (reusable templates) to MCP-aware AI clients. Servers can run locally (stdio transport) or remotely (HTTP/SSE transport).

Two Server Transports

TransportUse CaseExample
stdio (local)The server runs on your machine; the AI client launches it as a subprocessFilesystem MCP, local Postgres
HTTP / SSE (remote)The server runs on a vendor's infrastructure; the AI client connects over HTTPSunLocked CRM MCP, GitHub MCP

What an MCP Server Does (Lifecycle)

  1. Initialize — handshake with the AI client, exchange capabilities
  2. List tools — return a JSON manifest of every callable action
  3. Handle tool calls — execute the action and return structured results
  4. Stream responses — for long-running tasks, push partial results via SSE
  5. Manage auth — validate bearer tokens / OAuth scopes per tool

What a Tool Definition Looks Like

{
  "name": "search_contacts",
  "description": "Search insurance contacts by name, phone, email, or pipeline stage.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "query": { "type": "string" },
      "stage": { "type": "string", "enum": ["new", "qualified", "quoted", "bound"] }
    },
    "required": ["query"]
  }
}

Examples of MCP Servers in Production

  • unLocked CRM MCP — 16 insurance tools (contacts, quotes, policies, commissions, SOAs)
  • GitHub MCP — read repos, create issues, manage PRs
  • Linear MCP — create and triage issues
  • Stripe MCP — read payments, refund charges
  • Filesystem MCP — read/write local files (local stdio)

Security Model

Remote MCP servers should:

  • Authenticate every request (bearer token or OAuth 2.1)
  • Scope tokens to a tenant / workspace
  • Sign webhooks (HMAC-SHA256 with replay protection)
  • Rate-limit per token
  • Log every tool call for audit

How to Connect an MCP Server

For Claude Desktop, edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "unlocked-crm": {
      "url": "https://mcp.unlockedcrm.ai/v1/YOUR_TENANT",
      "headers": { "Authorization": "Bearer YOUR_TOKEN" }
    }
  }
}

MCP Server vs MCP Client

Server = the thing that exposes tools (your CRM). Client = the thing that calls them (Claude, ChatGPT, Cursor). One server can talk to many clients; one client can connect to many servers.

See unLocked CRM in Action

unLocked CRM ships a native MCP server, OpenAPI 3.1 spec, free A2P 10DLC registration, and remote SOA signing — built for the modern AI-first insurance agency. Book a 15-minute demo.

Frequently Asked Questions

Do I need to write my own MCP server?

Only if you're a tool provider. Most users just connect existing MCP servers (unLocked CRM, GitHub, Linear) to their AI client.

What language are MCP servers written in?

Any language. Anthropic publishes official SDKs in TypeScript, Python, Java, C#, and Kotlin.

Can one MCP server serve multiple users?

Yes — remote MCP servers are typically multi-tenant, with each tenant identified via a scoped URL or bearer token.