$ man how-to/what-are-mcps

Without MCP, your AI agent is trapped inside the code editor. It can read files, write code, and run terminal commands. That is it. It cannot see your Slack messages. It cannot pull leads from your CRM. It cannot push a draft to your newsletter platform. It cannot check your campaign analytics. Every time you need information from an external tool, you have to manually copy it and paste it into the chat. That is the problem. MCP is the solution. Model Context Protocol servers are bridges between your AI agent and your production tools. Each MCP server connects to one external tool and gives the agent a set of actions it can take. Read a Slack channel. Export HeyReach leads. Create a Substack draft. Deploy to Vercel. The agent can now reach outside the editor and interact with the real world.

PATTERN

An MCP server is a small process that runs in the background when your IDE starts. It connects to an external service using API credentials you provide. It exposes a set of tools (actions) that the AI agent can call. When you tell Claude "check the partner Slack channel," Claude sees the Slack MCP server in its tool list, calls the appropriate tool (read channel messages), and gets the results back. You never interact with the MCP server directly. You interact with Claude. Claude interacts with the MCP server. The MCP server interacts with Slack. Think of it as giving Claude hands. Without MCP, Claude can think and speak. With MCP, Claude can reach out and do things in the tools you already use.

CODE

Every MCP server follows the same setup pattern: 1. Get an API key from the external tool (Slack bot token, HeyReach API key, etc.) 2. Add a config entry to your MCP config file (.cursor/mcp.json for Cursor, ~/.claude/mcp.json for Claude Code) 3. The config specifies the server name, the command to start it, and the API key as an environment variable 4. Restart your editor. The server starts automatically. 5. Test it by asking Claude to perform an action with that tool. The config format is JSON. Each server gets a block with a command (usually npx), arguments (the npm package name), and environment variables (your API keys). You never commit this file to Git because it contains secrets. Add it to .gitignore.

PRO TIP

Do not install 15 MCP servers on day one. Each server adds startup time and consumes resources. Start with the two or three tools you use most. If you run outbound campaigns, start with your outreach tool (Instantly or HeyReach) and Slack. If you publish content, start with your publishing platform (Typefully or Substack) and your analytics tool. Build workflows around those first MCPs. Get comfortable with the pattern of telling Claude to interact with external tools. Then add more as your workflows demand it. My setup grew organically. I started with GitHub and Slack. Then added HeyReach. Then Substack. Then Browserbase. Each one was added because a specific workflow needed it, not because I was collecting servers.