Solution: MCP Server as a Bridge
MCP Server as a Bridge
The universal protocol that connects Claude to everything
Now that you understand the problem — Claude cannot see your live data — let us look at the solution. The Model Context Protocol (MCP) is an open standard created by Anthropic that gives AI models a universal way to connect to external systems.
Think of MCP as a USB-C port for AI. Just like USB-C gives you one standard connector for charging, data transfer, and video output, MCP gives Claude one standard protocol for reading data, executing actions, and using templates from any system.
The Architecture: Host, Client, Server
Host (Claude Code, Claude Desktop)
The application where you interact with Claude. It manages the conversation and coordinates everything. The host does NOT talk to MCP servers directly — it delegates to clients.
Client (Built into the Host)
A component inside the host that manages the connection to one MCP server. Each server gets its own client. It handles starting the server, sending requests, receiving responses, and lifecycle management.
Server (What You Will Build)
The server is what you create. It exposes capabilities to Claude through a standardized interface, connecting to your data sources — APIs, databases, files.
The Communication Flow
User asks Claude
You ask: "Show me today's new customers"
Claude finds the tool
Claude sees it has an MCP tool called get_customers
Host delegates to client
The host sends a request to the MCP client
Client calls the server
The client forwards the request to your MCP server
Server queries data
Your server queries the CRM database
Results flow back
Server to client to host to Claude to you
The Three MCP Primitives
Tools — Actions
Functions that Claude can call with input parameters and return results. Like API endpoints for Claude. Examples: search_customers, create_order, send_email.
Resources — Data
Data sources Claude can access, working like files with a URI. Typically application-controlled. Examples: config://app/settings, db://customers/recent.
Prompts — Templates
Pre-built conversation templates that help start common workflows with the right context. Examples: analyze_customer, weekly_report.
Most Common Primitive
Tools are the most commonly used primitive. They are model-controlled — Claude decides when and how to use them based on your request.
How This Connects to Plugins
| Plugin Component | Purpose | MCP Equivalent |
|---|---|---|
| SKILL.md | Instructions for Claude | Prompts (templates) |
| Commands | User-triggered workflows | Tools (Claude-triggered) |
| References | Static documentation | Resources (live data) |
| .mcp.json | Server configuration | Tells host which servers to start |
Plugin Only vs Plugin + MCP
Plugin Only
The skill contains instructions: "Use this script template to call the API." Claude generates a script, runs it, and the action completes. It works, but Claude cannot check data first without running another script.
Plugin + MCP Server
The MCP server exposes search_contacts and create_contact tools. Claude first searches for existing contacts, sees there is no duplicate, and then creates the new one — all in a single, smooth conversation.
Key Takeaway
An MCP server is a bridge between Claude and your systems. It turns your APIs, databases, and services into tools that Claude can use naturally — just like its built-in capabilities. In the next section, we will build one from scratch.
There are no comments for now.