MCP Server Setup: Connect AI to Your Business Tools

Right now, using AI for business means copying data from your CRM into ChatGPT, waiting for an answer, then copying it back. It is slow, error-prone, and breaks the moment your data changes.

MCP (Model Context Protocol) fixes this. It lets AI assistants query your tools directly — read customer records, update deals, check inventory, or run reports — all within the conversation. No copying. No context switching.

This guide shows you how to set up MCP servers step by step. We will start with a simple example and work up to connecting your actual business tools.

What is MCP and Why It Matters for Small Business

MCP is an open standard created by Anthropic (makers of Claude) that lets AI assistants connect to external data sources and tools. Think of it as a USB-C port for AI — one standard connector that works with any tool.

Before MCP, every AI-tool integration was custom. You needed a different script for Slack, another for Notion, another for your CRM. MCP replaces all of that with one protocol.

What You Can Do With MCP

• Query your database: "Show me all customers who bought in the last 30 days but haven't returned"
• Update your CRM: "Mark this lead as qualified and schedule a follow-up for Tuesday"
• Check inventory: "Do we have enough stock for next week's orders?"
• Generate reports: "Create a summary of this month's sales by product category"
• Draft emails from data: "Write a win-back email to customers who haven't purchased in 90 days"

The business value: Tasks that used to take 30 minutes of data gathering now take 30 seconds. Your AI assistant becomes a real team member with access to your systems.

MCP Architecture: How It Works

MCP has three components. Understanding them makes setup much easier.

How it flows: You ask Claude a question. Claude checks what MCP servers are connected. It sends a query to the relevant server. The server fetches data from your tool and returns it. Claude uses that data to answer you.

The magic: all of this happens automatically. You do not write API calls. You just ask questions in plain English.

Setting Up Your First MCP Server

We will start with the simplest possible setup: connecting Claude Desktop to a filesystem MCP server that lets Claude read and write files on your computer.

Prerequisites

• Node.js 18+ installed (node --version to check)
• Claude Desktop installed (Mac or Windows)
• A terminal or command prompt

Step 1: Install the MCP Filesystem Server

npm install -g @modelcontextprotocol/server-filesystem

Step 2: Configure Claude Desktop

Open Claude Desktop and go to Settings → Developer → Edit Config.

This opens a file called claude_desktop_config.json. Add this:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/path/to/your/documents"
      ]
    }
  }
}

Replace /path/to/your/documents with an actual folder on your computer. On Mac, this might be /Users/yourname/Documents. On Windows,C:\Users\YourName\Documents.

Step 3: Restart Claude Desktop

Quit and reopen Claude. You should see a hammer icon (🔨) in the chat input. This means MCP is active.

Step 4: Test It

Ask Claude:

List all files in my documents folder and tell me what each one is about.

Claude will query the filesystem server and read your files to answer. If it works, your first MCP server is live.

Connecting Claude Desktop to Multiple MCP Servers

Once you have the filesystem server working, adding more servers follows the same pattern. Claude Desktop can run multiple MCP servers simultaneously.

Example: Multiple Servers

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/yourname/Documents"]
    },
    "sqlite": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-sqlite", "/path/to/your/database.db"]
    },
    "fetch": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-fetch"]
    }
  }
}

With this config, Claude can: read your files, query your SQLite database, and fetch data from any URL — all in one conversation.

Where to Find MCP Servers

• Official Anthropic servers: filesystem, sqlite, fetch, puppeteer (browser automation)
• Community servers: GitHub, Slack, Notion, HubSpot, Stripe, Google Sheets
• npm search: npm search mcp-server finds community-built servers
• Build your own: MCP SDK supports TypeScript, Python, and Go

Connecting ChatGPT and OpenAI to MCP

ChatGPT does not natively support MCP yet (as of May 2026). But you can use MCP with OpenAI's API through frameworks like OpenAgents or by building a custom integration.

Option 1: OpenAgents (Easiest)

OpenAgents is an open-source framework that adds MCP support to OpenAI models. It runs locally and connects to the same MCP servers as Claude.

# Clone OpenAgents
git clone https://github.com/OpenBMB/OpenAgents.git
cd OpenAgents

# Install dependencies
pip install -r requirements.txt

# Configure your OpenAI API key
export OPENAI_API_KEY=your-key-here

# Start the server
python app.py

Option 2: Custom Integration (More Control)

If you are building a product or internal tool, you can use the MCP Python SDK directly with OpenAI's API.

from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client
import openai

# Connect to an MCP server
server_params = StdioServerParameters(
    command="npx",
    args=["-y", "@modelcontextprotocol/server-sqlite", "/path/to/db.db"]
)

async with stdio_client(server_params) as (read, write):
    async with ClientSession(read, write) as session:
        await session.initialize()
        
        # List available tools
        tools = await session.list_tools()
        
        # Query the database through AI
        response = await openai.chat.completions.create(
            model="gpt-4o",
            messages=[{"role": "user", "content": "Show me sales this month"}],
            tools=tools
        )

Note: Native MCP support in ChatGPT is expected in late 2026. Until then, Claude Desktop has the best MCP experience.

Database MCP Server: Query Your Data in Plain English

The SQLite MCP server is the easiest way to start. For production, you will want PostgreSQL or MySQL.

PostgreSQL MCP Server

npm install -g @modelcontextprotocol/server-postgres

{
  "mcpServers": {
    "postgres": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-postgres",
        "postgresql://user:password@localhost:5432/your_database"
      ]
    }
  }
}

What You Can Ask

• "How many orders did we get last week?"
• "Show me the top 10 customers by lifetime value"
• "Which products have inventory below 10 units?"
• "Create a report of monthly revenue for the last 6 months"

Security tip: Create a read-only database user for MCP. Never use your admin credentials. If MCP only needs to read data, restrict it to SELECT permissions.

CRM MCP Server: HubSpot Example

Connecting your CRM is where MCP gets really powerful for sales and customer service. Here is how to set up HubSpot (the pattern is similar for Salesforce and Pipedrive).

Step 1: Get Your HubSpot API Key

Go to HubSpot → Settings → Integrations → API Key. Copy your private app token.

Step 2: Install the HubSpot MCP Server

npm install -g @modelcontextprotocol/server-hubspot

Step 3: Configure

{
  "mcpServers": {
    "hubspot": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-hubspot",
        "--token",
        "your-hubspot-token-here"
      ]
    }
  }
}

What You Can Do

• "Show me all deals closing this month"
• "What was the last interaction with [customer name]?"
• "Create a new contact for [name, email, company]"
• "Which leads have not been contacted in 7 days?"
• "Summarize this deal's history for me"

For Pipedrive, Salesforce, Zoho: Community MCP servers exist for all major CRMs. Search npm search mcp-server-[crm-name] or check the MCP community repository.

Building a Custom MCP Server

If you have an internal API or a tool without a community server, building your own is straightforward. The MCP SDK handles the protocol; you just write the business logic.

Quick Start (TypeScript)

mkdir my-mcp-server
cd my-mcp-server
npm init -y
npm install @modelcontextprotocol/sdk zod

# Create src/index.ts

import { Server } from "@modelcontextprotocol/sdk/server/index.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { z } from "zod";

const server = new Server({
  name: "my-business-api",
  version: "1.0.0",
}, {
  capabilities: { tools: {} }
});

// Define a tool
server.setToolHandler("getCustomer", async (args) => {
  // Your API call here
  const customer = await fetch("https://api.mycompany.com/customers/" + args.id);
  return customer.json();
}, {
  id: z.string().describe("Customer ID")
});

// Start the server
const transport = new StdioServerTransport();
await server.connect(transport);

This is a minimal example. The full SDK supports resources (read-only data), tools (actions), and prompts (templates). See the MCP documentation for advanced patterns.

No-Code Alternative: n8n MCP Workflows

If the terminal setup feels intimidating, n8n (a visual workflow tool) now supports MCP through its AI agent nodes. No coding required.

How It Works

1. Install n8n (cloud or self-hosted)
2. Add an AI Agent node to your workflow
3. Connect MCP tools: HTTP Request, database query, CRM API
4. The AI agent automatically decides which tool to use based on your question

When to Use n8n vs Direct MCP

n8n has a free tier for up to 100 workflow executions/month. For small businesses, this is usually enough to start.

Security Best Practices for MCP

MCP servers have access to your business data. Follow these rules to keep things secure.

The MCP Security Checklist

• Use read-only database users. MCP rarely needs write access. If it does, use a dedicated user with minimal permissions.
• Store API keys in environment variables. Never paste secrets into config files that might be committed to git.
• Run MCP servers locally or on a private server. Avoid cloud-hosted MCP servers for sensitive data unless you trust the provider.
• Audit what MCP can access. Test with: "List all tools you have access to and what they can do."
• Monitor API usage. Unexpected spikes in API calls could indicate a loop or misuse.
• Keep MCP servers updated. Community servers may have security patches. Update monthly.

Red flag: If an MCP server asks for admin credentials to your CRM, database, or cloud infrastructure, say no. Build a limited-access account instead.

Troubleshooting Common MCP Issues

Issue: Claude Does Not Show the Hammer Icon

Fix: Check your config file syntax. Common causes:

• Missing comma between server entries
• Incorrect path format (Windows paths need double backslashes)
• The server package is not installed globally (try npm install -g)

Issue: MCP Server Crashes

Fix:

• Check that the tool/API credentials are valid
• Verify the connection string or API endpoint is correct
• Look at the server logs — most MCP servers output errors to stderr

Issue: Claude Cannot Find the Right Tool

Fix: Be specific in your prompt. Instead of "check my database," say "query my postgres database for customers who signed up last week."

Issue: Slow Responses

Fix: MCP adds latency because the AI makes multiple round-trips. For faster responses:

• Use local MCP servers (not remote APIs)
• Limit the data returned (add date filters or row limits)
• Use Claude 3.5 Sonnet instead of Opus for tool use — it is faster