What an MCP server is and how it connects AI to your tools

What the Model Context Protocol is

Anthropic introduced the Model Context Protocol on 25 November 2024 as «a new standard for connecting AI assistants to the systems where data lives». In December 2025 it donated it to the Agentic AI Foundation, a directed fund under the Linux Foundation.

Before a standard like this, every connection between an assistant and a tool had to be built separately. As Anthropic puts it, «every new data source requires its own custom implementation». With MCP, a tool is connected once and any compatible assistant can use it.

Servers and clients

The protocol has two main pieces:

• The MCP server is the program that offers a tool's data and actions: reading tasks, searching for an email, creating an appointment.
• The MCP client is the AI application that connects to it, for example an assistant like Claude or ChatGPT.

When you ask your assistant a question («what have I got on tomorrow?»), the assistant asks the tool's MCP server for the information, the server replies with the data you have permission to see, and the assistant answers you.

What it's for, day to day

If you're not a developer, you notice MCP in one way: you can ask your AI assistant about data that lives in other tools without copying and pasting it. Some examples:

• Ask the assistant for a summary of a client's outstanding tasks.
• Ask which meetings you have this week and what you need to prepare for them.
• Find an old email from a description, without remembering who sent it.

The assistant doesn't keep a copy of your tools: it queries them when needed, with the permissions you've given it.

How it differs from a traditional integration

Many tools already connect to each other: your calendar with your email, your task manager with your chat. These are custom integrations between two specific apps, and each one has to be built and maintained separately.

MCP differs in two ways. First, it's a shared standard: an MCP server works with any compatible assistant, not just one. Second, it's designed so that whatever uses the data is an AI assistant, which decides what information it needs to answer your question, instead of a fixed rule like «when A happens, do B».

That makes it more flexible, but it also makes choosing the right permissions more important: the assistant may ask for things you hadn't anticipated.

A step-by-step example

Imagine your tasks are in a tool with an MCP server and you use a compatible AI assistant:

1. In the assistant, you add the tool's MCP server and accept its permissions (for example, read-only).
2. You ask: «What have I got outstanding this week for client X?».
3. The assistant asks the server for that client's tasks dated this week.
4. The server returns only the ones you can see, and the assistant summarises them and suggests where to start.

If you stop using the assistant tomorrow, you revoke access and the server stops answering it. Your data stays in the original tool.

Common mistakes

• Granting write permission by default when you only need to look things up.
• Connecting third-party servers without checking who maintains them: an unofficial server may ask for more access than it needs.
• Forgetting about old connections. Once a quarter, review which assistants have access to which tools and revoke whatever you no longer use.

What to check before connecting one

Connecting an MCP server means giving an AI assistant access to a tool. Before you do, it's worth asking yourself four questions:

1. What can it read? All of the tool's data or just part of it.
2. What can it write? Whether it can create, change, send or delete things, and whether it asks you to confirm.
3. How is it revoked? Whether you can cut off access overnight, and from where.
4. Who maintains it? Whether it's the tool's official server or a third-party one.

A good rule of thumb to start with: read-only by default and write access only when you need it.

MCP and privacy

An MCP server doesn't change who sees your data in the original tool, but it does add a new reader: the AI assistant. What the assistant reads may pass through the AI provider you use, under its terms.

That's why it makes sense to connect only the tools you need, review permissions and revoke access for the ones you no longer use.

How duokim does it

duokim acts as an MCP server so you can query it from Claude or ChatGPT with your permission. It's a feature in testing in the beta:

• Read-only by default; writing, only if you turn it on.
• Never delete or send anything.
• Up to 3 connections at once, and you can revoke them whenever you want.

Inside duokim, Kim keeps suggesting and you keep deciding; your AI assistant looks up what it needs there.

Claude and ChatGPT + duokim (MCP server)

Frequently asked questions

Sources

• Anthropic · Introducing the Model Context Protocol (accessed 11 November 2026)
• Anthropic · Donating MCP to the Agentic AI Foundation (accessed 11 November 2026)

Article by the duokim team. If you spot a mistake, write to us and we'll correct it.