Stop using Claude as just a chatbot—MCP changes everything

AI chatbots are great at answering questions or analyzing text, but they're usually limited to the tools that are built in. MCP changes that by giving chatbots a standard way to connect to external services, so that you can use your chatbot to actually get things done.

What is MCP, and why does it matter?

A universal connector for your chatbot

MCP stands for Model Context Protocol. It's an open standard that was introduced by Anthropic, the creators of the Claude chatbot . It's designed as a standardized way to connect external services to your chatbot.

Before the introduction of MCP, it was possible to connect your AI chatbot to external services using custom connectors, but these would usually only work with one specific service and were hard to reuse or maintain. MCP allows any compatible AI client to talk to any compatible external service using a common framework, making it easier to connect services to AI chatbots.

Although the standard was developed by Anthropic, it's been adopted by other platforms, including ChatGPT and Google. Using MCP, your AI chatbot can go from being a text-based chatbot to a working assistant that can perform actions on your behalf.

What you can do with Claude and MCP

Use Claude to take actions

Anthropic introduced MCP, so it's no surprise that Claude is one of the most accessible ways to use connected services. There is a huge number of MCP connectors you can use in Claude, which allow you to perform actions outside of the chatbot itself.

There's a Filesystem connector that lets Claude read, write, and create files and folders on your local computer, a GitHub connector that you can use to let Claude manage pull requests or search through code, a Notion connector that can read pages or add new database entries , and even a Slack connector that you can use to send Slack messages directly from Claude.

If an MCP connector for a service doesn't exist, in some cases, it's possible to get Claude to create one for you and run your own MCP server to connect to that service. While this isn't always possible, it can be a great way to integrate even more services with Claude.

For example, I'm running the Qwen3-TTS text-to-speech model on my mini PC to generate spoken audio announcements for my smart speakers . I wanted to be able to generate speech directly from Claude so that I could do things such as generating multiple audio files of the same messages worded in different ways, rather than having the exact same response each time. Using Claude, I created a custom MCP connector that lets me generate speech directly from within Claude chats.

How to set up an MCP in Claude

Connecting your first MCP server

Setting up some of the popular MCPs in Claude is simple to do. There are official MCPs called connectors for services such as Google Calendar , Slack, GitHub, and more. You can connect these directly in Claude without any additional configuration.

In a Claude chat, click the "+" (plus) icon. Hover over "Connectors," and click "Manage connectors." Click the "+" (plus) icon and select "Browse connectors." You can now choose and install any of the available connectors, and Claude will configure everything for you. Alternatively, select "Add custom connector," and add the name and URL of your remote MCP server.

To add a local custom MCP, you need to add the server details to yourclaude_desktop_config.jsonfile. This is located in~/Library/Application Support/Claude/on macOS, and%APPDATA%\\Claude\\on Windows. You add an entry under themcpServerskey, in the following format:

Command is the program that runs your code, which could bePython,node,uv, or something else.

Finally, restart Claude, and you should see your MCP listed with any other connectors when you click the "+" (plus) icon in the chat window.

The limitations of MCP

Claude is less limited than ChatGPT

While MCP can make your AI chatbot far more useful, it's not perfect. There's a lot you still can't do, and it's not without its risks. MCP servers can read and write real data, and since chatbots can hallucinate, there's a risk that you could lose important data when using an MCP.

Not every MCP works that well or has the access you might need. The official connectors are solid, but community MCP servers aren't always completely reliable or well-maintained. There's also a risk of prompt injection , where malicious content on a web page or in a file could trick the AI into misusing the connected tool. Anthropic has steps in place to mitigate this risk.

It's important to keep these risks in mind when you're using an MCP. If you limit what an MCP can access and don't use something you're not confident in, however, then you should be able to use many MCPs safely.

MCP capabilities will differ depending on the chatbot you use, too. For example, currently ChatGPT's support is for remote MCP servers, whereas Claude will also let you use local MCP setups.

---

Get more from your chatbot

AI chatbots are impressive, but they're mostly stuck in their own walled gardens. With MCP, you can let your AI talk to other platforms and services, and make it something far more useful.