How to Set Up and Use an MCP Server
MCP setup is technically feasible for anyone, but it's worth being honest about what's involved. You're editing JSON configuration files, working with terminal commands, and navigating concepts like API authentication and server transports. For marketing ops professionals who spend their days in Marketo and Salesforce, this is unfamiliar territory.
The good news is that the AI tools you're already using can help. Claude Code, for instance, is competent enough that once you tell it which MCP server you want to connect, it can handle the configuration settings for you. The process itself comes down to adding a few lines to a config file and restarting your application, but knowing that doesn't make the first time through less intimidating, and there's no shame in asking your AI assistant for help getting set up. Most people building AI workflows today, including experienced operators, have leaned on their AI tools to handle MCP configuration at some point.
How MCP servers work
Before the setup steps, it helps to understand what you're actually configuring.
An MCP server is a program that exposes tools and data to AI assistants using a standardized protocol. MCP servers are downloaded from public registries on the internet and then run locally on your machine, providing your computer with scripts it can execute when your AI assistant needs to access external data. The connection between your AI client and the MCP server is handled via API calls, which means MCP is, in many ways, an advanced API with additional documentation and workflows layered on top. It gives your AI assistant not just access to data, but a road map for how to interact with the product, the vendor, and the service.
When you connect an MCP server, your AI assistant gains new capabilities. Connect a filesystem server and your assistant can read files on your computer. Connect a GitHub server and it can browse your repositories. Connect a production platform's server and it can query your campaigns, brands, and templates through the same infrastructure that governs the human workflow.
Two types of MCP servers exist, and they set up differently:
Type | How it runs | What you configure | Best for |
|---|---|---|---|
Local (stdio) | Your AI client starts it as a background process on your machine | A command and arguments in a config file | Personal tools, file access, local data |
Remote (HTTP) | Runs on a server hosted by the vendor, you connect via URL | A URL and authentication credentials | SaaS products, team tools, shared platforms |
Type | Local (stdio) |
|---|---|
How it runs | Your AI client starts it as a background process on your machine |
What you configure | A command and arguments in a config file |
Best for | Personal tools, file access, local data |
Type | Remote (HTTP) |
|---|---|
How it runs | Runs on a server hosted by the vendor, you connect via URL |
What you configure | A URL and authentication credentials |
Best for | SaaS products, team tools, shared platforms |
Local servers are more common for getting started. You add a command to your config file and your AI client manages the rest. Remote servers require a URL and usually OAuth authentication, similar to connecting any web application where you sign in with your credentials.
Setting up MCP in Claude Desktop
Claude Desktop was the first AI client to support MCP, and its setup process is the most established.
Step 1: Find the config file
Open Claude Desktop and go to Settings, then look for the Developer section (some versions label it Extensions). Click "Edit Config" and it opens the JSON file directly. If you prefer to navigate manually:
- Mac:
~/Library/Application Support/Claude/claude_desktop_config.json - Windows:
%APPDATA%\Claude\claude_desktop_config.json
If the file doesn't exist yet, create it.
Step 2: Add a server
Start with the filesystem server. It's the simplest MCP server available, built by Anthropic as a reference implementation. It requires no API keys and gives you an immediate result to verify that everything works.
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/yourname/Desktop"]
}
}
}
Replace /Users/yourname/Desktop with any folder path you want your AI assistant to access. Only directories you explicitly list are accessible, so the server is sandboxed to what you specify.
Step 3: Restart Claude Desktop
Close and fully quit the application (Cmd+Q on Mac, not just closing the window). This is the most common point where people get stuck: closing the window doesn't restart the application, and the MCP server won't connect until Claude Desktop fully relaunches.
Step 4: Verify the connection
In a new chat, click the tools icon near the input box. Your filesystem server should appear in the list of connected tools. Ask Claude something like "list the files on my Desktop" and it will use the MCP server to read your filesystem.
Leah Miranda, Head of Demand Gen & Lifecycle at Zapier, describes what the broader potential feels like in practice: "You're at a conference, you snap a photo of someone's badge and ask 'Is she a customer?' MCP checks your systems, pulls the right data, and returns a clear summary of who she is, her account status, and how she's using your product." The filesystem example is smaller than that vision, but the principle is identical. Your AI assistant is reaching outside its sandbox and accessing something real.
Troubleshooting common issues
If the server doesn't connect after restarting:
- Validate your JSON at jsonlint.com for syntax errors. Trailing commas after the last item in an object or array are the most common issue and they fail silently, meaning you won't get an error message explaining what went wrong.
- Use absolute file paths, not relative ones.
/Users/yourname/Desktopworks consistently.~/Desktopsometimes doesn't, because Claude Desktop may launch servers from the system root directory rather than your home folder. - Check the logs at
~/Library/Logs/Claude/mcp*.logon Mac. These show the exact error messages from the MCP server startup process and are usually the fastest way to diagnose connection issues. - Environment variables aren't inherited from your shell profile. If a server needs tools installed via npm or Homebrew, you may need to specify the full path to the executable in your config rather than relying on your shell's PATH variable.
Setting up MCP in ChatGPT
ChatGPT handles MCP differently from Claude Desktop. It only supports remote MCP servers, meaning you can't connect local tools running on your machine. You connect to servers that are hosted by the vendor on the internet.
For marketing ops, this is actually the more relevant model for SaaS products. CRMs, marketing automation platforms, and analytics tools deploy remote MCP servers that you authenticate against with your existing credentials.
How to connect
- In ChatGPT, go to Settings, then Apps and Connectors
- Enter the URL of the remote MCP server provided by the vendor
- Authenticate through the OAuth flow that opens in your browser (similar to "Sign in with Google")
- Once connected, interact with the server's tools directly in conversation
The key difference from Claude Desktop: ChatGPT requires remote servers only. If you want to test with a simple local server like the filesystem example above, Claude Desktop or Cursor are better starting points. ChatGPT's MCP support is best suited for connecting to SaaS platforms that have deployed their own remote MCP servers.
Setting up MCP in Cursor
Cursor uses the same JSON config format as Claude Desktop, which makes moving between the two straightforward if you're already familiar with one.
Config file locations
- For a specific project:
.cursor/mcp.jsonin your project root - For all projects:
~/.cursor/mcp.jsonin your home directory
Local server config
The format is identical to Claude Desktop:
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/yourname/Projects"]
}
}
}
Remote server config
For connecting to hosted MCP servers:
{
"mcpServers": {
"remote-tool": {
"url": "https://app.example.com/mcp/public"
}
}
}
Cursor also supports one-click installs from its marketplace and cursor.directory, where you can browse available MCP servers and install them without editing JSON manually. For debugging, open the Output panel (Cmd+Shift+U) and select "MCP Logs" from the dropdown to see connection status and specific error messages.
What to try after your first connection
Once the setup works with the filesystem server, the natural next step is connecting MCP servers that match your actual workflow. Each additional server follows the same configuration pattern, so the learning curve flattens quickly after the first successful connection.
Server | What it connects | Why it's useful for marketing ops |
|---|---|---|
GitHub | Repositories, issues, PRs | Track campaign project work from your AI assistant |
Slack | Channels, messages, threads | Search conversations and extract action items without switching apps |
Google Drive | Documents, spreadsheets | Pull data from shared files into AI conversations |
Notion | Databases, pages, properties | Query project management data and campaign planning docs |
Server | GitHub |
|---|---|
What it connects | Repositories, issues, PRs |
Why it's useful for marketing ops | Track campaign project work from your AI assistant |
Server | Slack |
|---|---|
What it connects | Channels, messages, threads |
Why it's useful for marketing ops | Search conversations and extract action items without switching apps |
Server | Google Drive |
|---|---|
What it connects | Documents, spreadsheets |
Why it's useful for marketing ops | Pull data from shared files into AI conversations |
Server | Notion |
|---|---|
What it connects | Databases, pages, properties |
Why it's useful for marketing ops | Query project management data and campaign planning docs |
The MCP ecosystem includes over 10,000 public servers, with registries like Smithery, PulseMCP, and the official MCP Registry cataloging them by category. Marketing-specific servers for CRM, analytics, and advertising platforms are already available, and the list grows regularly.
MCP servers are complementary by design. Having connections to your CRM, your analytics platform, your project management tool, and your email production platform gives your AI assistant access to the full workflow context rather than isolated pieces. Each server you add extends what your AI operator workflow can accomplish, giving the AI additional context and additional tools to act on it.
Explore how Knak's AI features connect to your brand context.
