🚀 MCP Architecture Explained: Connecting Claude AI to Salesforce

💬 In plain words: The Model Context Protocol (MCP) is an open standard created by Anthropic that acts like a universal plug for AI. Instead of writing custom code to connect an AI (like Claude) to every single database or app, MCP provides a structured, secure way for the AI to ask enterprise systems (like Salesforce) for data and actions.

⚡ Key Points: 1-Minute Summary

• Standardized Connections: MCP replaces messy, custom API integrations with a unified client-server architecture for AI.
• The Host: The application where the user talks to the AI (e.g., Claude Desktop).
• The Client: The communication layer bridging the AI and the data.
• The Server: The application that securely exposes your enterprise data (e.g., a Salesforce environment).
• Capabilities: Servers provide the AI with Resources (read-only data), Tools (executable actions), and Prompts (reusable instructions).

🗺️ Module Map: Claude Foundation

CLAUDE FOUNDATION
├── What Is Claude?
├── Claude Model Families
├── Claude Reasoning
├── Claude Tool Use
├── AI Agents
├── Model Context Protocol (MCP)
├── MCP Architecture ← Current Topic
└── Claudeforce Architecture

🎬 Real-Life Example: The Ultimate AI Employee

Scenario: Preparing for a Customer Meeting

A sales manager tells their AI agent: "Prepare me for tomorrow's customer meeting."

To do this well, the AI needs data scattered across multiple places: opportunities in Salesforce, support tickets in Zendesk, and meeting times in Google Calendar.

Instead of building three custom integrations, the AI Host uses its MCP Client to instantly talk to the MCP Servers running on Salesforce, Zendesk, and Google. The servers expose the approved capabilities, the AI collects the relevant context, and it generates a perfect briefing document in seconds.

🧩 Core Concepts: Breaking Down the Architecture

To understand how MCP securely streams data to LLMs, you need to know its three main architectural layers:

• The Host: This is the AI application where the user interacts. Think of the Claude Desktop app or an enterprise AI portal.
• The Client: Sitting inside the Host, the MCP Client is the actual engine managing the 1:1 communication protocol with external servers.
• The Server: A lightweight program that connects directly to a data source (like Salesforce) and translates its specific data into the universal MCP format.

When an MCP Server connects to a Client, it offers three primary capabilities:

• Resources: Information the AI can read (e.g., A customer's Account Record or an API documentation file).
• Tools: Actions the AI can take on the user's behalf (e.g., "Create a new Task in Salesforce" or "Update Opportunity Stage").
• Prompts: Pre-built templates to guide the AI's behavior consistently.

🔄 The MCP Communication Flow

User inputs a request (e.g., "Summarize Acme Corp")
      ↓
AI Host processes the intent
      ↓
MCP Client reaches out to available servers
      ↓
MCP Server (Salesforce) receives the request
      ↓
Server executes a Tool or retrieves a Resource
      ↓
Server returns the structured data to the Client
      ↓
Claude generates the final, context-rich response

⚖️ Component Comparison

• Host: The AI Application (Example: Claude Desktop App)
• Client: The Communication Manager (Example: Built-in MCP Connector)
• Server: The Capability Provider (Example: Salesforce MCP Server)
• Tool: Performs an Action (Example: Create a Lead)
• Resource: Provides Information (Example: Customer Support History)

🧭 360 Card — MCP Architecture

Rule: Each component (Host, Client, Server) has a strictly separated responsibility. The AI never touches the raw database directly.
Gain: Highly reusable AI integrations. You write an MCP Server once, and any MCP-compatible AI can use it.
Price: Requires careful architectural design, server hosting, and API monitoring.
Limits: MCP does not replace authentication or authorization. Security permissions are still governed by the underlying enterprise system (like Salesforce Profiles/Permission Sets).

⚠ INTERVIEW TRAP

Wrong: "MCP is a direct database connection that lets Claude query our SQL tables."
Correct: "MCP is a standard communication protocol. It allows AI clients to connect to secure MCP Servers, which expose highly controlled tools and resources, keeping the actual database safely abstracted away."

💬 Core Q&A & Follow-ups

Q: What is the exact difference between an MCP Client and an MCP Server?

A: The Client lives inside the AI application (the Host) and initiates requests. The Server lives closer to your business data, listens for those requests, securely interacts with the underlying system (like Salesforce), and returns the capabilities (Tools and Resources) to the Client.

Q: Can one AI Host connect to multiple MCP Servers at the same time?

A: Yes! This is the superpower of MCP. A single Claude application can connect to a Salesforce MCP Server, a local file system MCP Server, and a Slack MCP Server simultaneously, blending data from all three into one intelligent answer.

Q: Does the MCP Server decide who gets access to the data?

A: No. MCP is just a messaging protocol. Authentication, user authorization, and data governance remain the job of the enterprise systems. If a user asks Claude to fetch a Salesforce record they don't have permission to see in Salesforce, the MCP Server will return an access denied error.

Q: Follow-up: How do you build an MCP Server for Salesforce?

🎯 Deep Analysis: You can build custom MCP Servers using Anthropic's open-source SDKs (available in TypeScript and Python). For Salesforce, you would build a Node.js or Python server that uses the Salesforce REST API or SOQL to fetch data, and then wraps those functions into MCP Tools and Resources that Claude can understand.

📌 The Ultimate Takeaway

Claude provides the Intelligence.
MCP provides the Connection Standard.
The MCP Server provides the Capabilities.
Salesforce provides the Business Data.
The AI Agent executes the Goal.

See also: The Ultimate Guide to Model Context Protocol (MCP): The USB-C for AI