The Protocol Behind the Curtain

MCP servers are quietly connecting AI agents to everything, your email, calendar, files, and more. Here's what they are and what you need to know before you trust one.

Pull back the curtain and the connections are already running.

MCP (Model Context Protocol) servers have quietly become the backbone of how AI agents operate. Here's what they are, why they matter, and what you need to know before you trust one.

How MCP Simplifies AI Integration

Imagine you have a bunch of helpful experts sitting in separate rooms. One knows your Google Calendar inside and out, one lives in your inbox, one can read every file on your computer. Normally, getting help from all of them means knocking on each door individually, speaking a different language every time.

MCP is the universal hallway that connects all those rooms and a common language everyone already speaks.

Before MCP, connecting an AI agent to five different tools like Google Calendar, Slack, a file system, a database, and your inbox meant building five separate custom integrations. Each one built differently, maintained separately, and creating its own security considerations.

MCP changes that. It is an open standard, developed by Anthropic (Claude creator) and now widely adopted, that lets AI agents connect to any tool that has an MCP server built for it, using the same protocol every time. One standard. Any tool. Plug and play.

That is genuinely useful. It is also why understanding MCP matters from a security standpoint.

🚀 MARTY SAYS

"A hallway connects every room, but it also means anyone in the hallway has access to all of them. MCP is no different. Convenience and access always travel together."

Why MCP Matters for How You Work

MCP servers are already built for hundreds of tools like Gmail, GitHub, Slack, databases, internal APIs, and more. When you give an AI agent access to an MCP server, you are giving it the ability to read from and write to whatever that server connects to.

That is powerful. Depending on what you connect, an agent can schedule and cancel meetings, read and send emails, or create, move, and delete files, all on your behalf.

The agent is only as limited as the permissions you grant the MCP server. And that is exactly where the security conversation starts.

THE PRACTICAL PICTURE

MCP is to AI agents what app permissions are to your smartphone. When an app asks for access to your contacts, location, and microphone, you make a decision about whether that's necessary for what the app does. MCP access works the same way. The question is always: does this agent actually need this level of access to do its job?

The Security Baseline for MCP

Least privilege. Always. The access you grant is the access that can be exploited. Give every agent the minimum it needs for the specific task and nothing more.

Know what's connected. Every active MCP connection is an access point. If you don't know what's connected, you don't know your exposure.

Trust the server, not just the agent. An MCP server is a piece of software. It can have vulnerabilities, be misconfigured, or be compromised. Know who built it and what data it touches before you enable it.

Know who built the server. An MCP server is third-party software. Before you enable one, ask whether it comes from a reputable source and whether it has been vetted by your organization.

Safe Harbor: Three Things You Can Do This Week

Audit your active MCP connections. If you use any AI tool with agent capabilities, find the integrations or connected apps settings. List everything that's enabled. Remove anything you don't actively need.
Apply the Need-to-Know Rule to every MCP permission. For each connection, ask: does this agent need read access, write access, or both? Can you limit it to read-only for now?
Default to read-only. When in doubt about whether an agent needs write access, start with read-only. You can always expand permissions later. Taking them back is harder.

Next week: We zoom out. Where is all of this actually headed in 2026 and what are the four shifts that matter most for the people reading this?