As much as I like to hate Microsoft's Copilot, I do think there was some potential initially for it to be good, especially with its Windows integration. Being able to control parts of your PC using AI was kind of cool, as you no longer had to manually navigate pages and settings to get where you wanted. But of course, as with all interesting ideas, Microsoft gave up on it and turned Copilot into just another AI assistant.

Thankfully, Windows-MCP is a third-party tool that's here to pick up the slack and offer the AI features Microsoft won't. Relying on other LLMs such as Anthropic's Claude, Windows-MCP can take control of your PC in various ways, opening apps and websites, changing settings, and even installing apps. It's exactly what I wish Microsoft had done with Copilot (and later Copilot+) should have been — even if the actual implementation isn't that good.

It takes some work to set up

Be prepared

Right off the bat, setting up Windows-MCP is a bit of a pain, but if you follow all the steps correctly, it should be fine. The GitHub project lists a few requirements, including Python 3.13 or newer (at writing time, this is the latest version), the Claude desktop app from Anthropic, the UV package manager for Python, and the Desktop Extension CLI tool from Anthropic.

A couple of requirements aren't mentioned, though. You'll need to have Node.js installed, preferably with NVM set up, in order to install the Desktop Extension tool, and creating the extension itself requires you to clone the GitHub repository for Windows-MCP, which is done with a Git command, meaning you may also want to install Git. Of course, you could also download the repository and extract it manually, but if you're following the recommended steps, there are a few things you need.

It's not the end of the world, but it does take some time to get things going. Once you have everything installed, you need to create the Windows-MCP extension for Claude, and then load it in the Claude desktop app. At this point, you're ready to use it.

Controlling your PC with text commands

Copilot should have done this

Once you go through the mildly tiring setup process, you can use Claude to control your PC with various text commands. The process is fairly transparent, so when you issue a command, you can see every action Claude is taking in order to achieve the result you want, whether it's checking the current state of your machine, launching an app, clicking a button, or entering text. Claude needs your permission to perform each different kind of action, so you have a layer of security, though responding to the permission prompts can sometimes break the flow, so it's easier to always allow each action so it can do its work.

I'm on the free plan of Claude, so I'm limited in how many interactions I can make, but I've played around with it a bit. I've had it open and close my browser, whether that's Vivaldi or Edge, open the Calculator app, and even open specific websites. I specifically asked it to open XDA in Vivaldi, and I got to watch the process of opening the browser, selecting the address bar, and entering the URL. Claude even provided a quick summary of the content on the page.

Since it relies on an LLM, there's a lot more you can do with specific interactions just by using your language, but again, I'm on the free plan, so I only got to try a few things.

Is it worth it, though?

Who needs this?

As interesting as this idea is, I do have to wonder if there's a reason for most people to use it. Claude's responses and taking each of the steps in a process takes significantly longer than doing it yourself. I'm sure that's to be expected to some extent, but it's still unfortunate that the difference is so big. I had to just sit and stare at the screen for what felt like multiple minutes while Claude figured out how to open Vivaldi and then type the XDA URL into the address bar, and the entire time, I kind of wished I would just do it myself.

I can see this being useful for things you might not know how to do, like changing the display resolution, or turning off notifications, but that's about it.

Then there's the fact that the experience is just buggy. I've had a few issues using Windows-MCP, starting with the fact that, somehow, when it opens Vivaldi, it completely ignores all the user data I already have in Vivaldi. My user profile, email accounts, settings, history — everything is just gone. I tried looking into it to see if it somehow installed a separate instance of Vivaldi, but that doesn't seem to be the case, and they're running from the same folder. I frankly don't even know how it's possible.

At one point, I've also tried to ask Claude to open XDA in Vivaldi twice in a row, and I closed the browser after the first command. However, when I made the second command, Claude still assumed Vivaldi was open and was trying to interact with it. Thankfully, it eventually realized the browser wasn't there and repeated the steps.

Between the sluggishness of the experience and the occasional errors, I do find that this isn't an experience that's really worth it. With a paid Claude plan, at least you'd have unlimited interactions and you'd be able to use it more seamlessly, but even then, I don't think this would make for a great experience, certainly not one I'd be willing to pay monthly for.

At least it's trying something interesting

While the experience of using Windows-MCP isn't particularly pleasant, I still find it more appealing than using Copilot, and if anything, it's a lesson to Microsoft, showing what could and should have been done with Copilot. An experience like this that's more integrated into the OS and more optimized for faster performance would make all the sense in the world, but instead, we're stuck with a glorified chatbot that can search the web and make soulless images.

I do hope Windows-MCP can continue to improve and become a more useful tool over time. As it stands, I'm not interested in using it, but at least the concept is more interesting than most of the things Copilot+ can do right now.