Skip to Main Content
PCMag editors select and review products independently. If you buy through affiliate links, we may earn commissions, which help support our testing.

Microsoft Thinks OpenClaw Is the Future of Windows. My Testing Says Otherwise

I spent days wrestling with Microsoft's experimental companion app, only to find a maze of terminals, rising API costs, and a tool that leaves regular users completely behind.

Senior Writer, Software
OUR EXPERT
Chris has been writing about Windows, PCs, and software for more than 15 years. His experience with PCs dates back to MS-DOS and Windows 3.1.
Edited By:  
August 6, 2026
Add as a preferred source on Google
An OpenClaw logo over a Windows 11 desktop background (Credit: Zain bin Awais/PCMag/OpenClaw)

Getting your Trinity Audio player ready...

Microsoft first showed off the native OpenClaw companion app for Windows during its Build keynote in June as proof that the future of the OS is agentic. But, based on my testing, the experimental, open-source AI tool still isn't ready for the limelight.

Even if Microsoft's dedicated app eases some of their setup complications, OpenClaw agents remain complex and nonintuitive. That's to say nothing of the fact that agents are increasingly expensive to run, and that both Claude and ChatGPT can also now interact with your PC fairly seamlessly. Meanwhile, Apple's macOS Golden Gate continues to integrate AI in practical, user-friendly ways, such as with the new Siri and vibe-coded browser extensions in Safari.

I understand that Microsoft is likely targeting developers and experienced AI enthusiasts with its OpenClaw app, but if it ever intends for regular users to get on board, the company has a lot of work to do. I don't think OpenClaw is anywhere near ready for the masses, but you can read about my full experience below.

OpenClaw Creator Peter Steinberger Talks Claw for Windows
PCMag Logo
OpenClaw Creator Peter Steinberger Talks Claw for Windows

Setup Process: An Easy Install That Quickly Turns Technical

The OpenClaw Windows app is a graphical installer that automatically sets up OpenClaw. The underlying OpenClaw software is still Linux-based, so it runs via the Windows Subsystem for Linux (WSL).

OpenClaw setup on Windows
The companion app sets up the underlying OpenClaw software with minimal user input (Credit: Microsoft/PCMag)

However, the installation process quickly becomes technical. You must set up a model provider, for example. Although you can use a ChatGPT subscription, most options require you to connect an API key and pay per token. I used OpenRouter, an API service that provides access to a wide variety of models. You can also run large language models locally with LM Studio, but this requires powerful hardware. From the start, it's clear that the OpenClaw app is primarily for AI enthusiasts and developers familiar with API endpoints, model choices, and the low-level details that consumer chatbots hide, not for regular Windows users.

You can use the chatbox in the app to send prompts to your agent. To send text messages to your agent when you’re away from your PC, you need to enable at least one chat channel. The recommended Telegram bot was simple to configure, though I had to open a terminal window and paste a command during setup. The Windows app doesn’t fully replace the need to use the terminal for OpenClaw in general, something less technical users will likely prefer to avoid. I can't stress how different this approach is versus Apple's with Apple Intelligence features throughout macOS Golden Gate. Apple's AI tools simply work without much configuration and live in approachable context menus and interfaces. Even ChatGPT and Claude's PC-focused agentic features are much easier to set up and use.

OpenClaw channel setup on Windows
The companion app requires extra legwork to set up a chat gateway and model provider (Credit: Microsoft/PCMag)

Microsoft's Agent Sandbox Keeps Your Files Safe from Your AI

On Windows, OpenClaw uses Microsoft’s new Microsoft Execution Containers (MXC) sandbox technology, which gives you granular control over what the agent can access. You can restrict access to specific folders and specify whether the agent can capture your screen or view your webcam. The sandbox aims to prevent AI agents from misbehaving and deleting all the files on your PC, for example. The app defaults to a locked-down security level, but I switched it to a recommended security level so the agent could access the internet and view files in my Downloads folder.

OpenClaw sandbox configuration on Windows
The OpenClaw Windows Companion app said the sandbox protection would be stronger if I were running an Insider build of Windows 11 on my PC (Credit: Microsoft/PCMag)

"You're Becoming Someone": Inside OpenClaw's Bizarre World

The OpenClaw Windows Companion app has a main window with a central chatbox and various settings panes. The app appears straightforward at first glance, but the complex assortment of options scattered across multiple panes makes things confusing. You can minimize the app to a tray icon if you prefer.

For testing, I used a US-hosted version of the DeepSeek V4 Flash model, which is less expensive than Anthropic's and OpenAI’s models. Microsoft might reportedly use this model for Copilot to help reduce corporate AI bills.

To start, the agent asked for an identity, vibe, and emoji. I let it pick its own and ended up with an agent called Nova who liked the sparkle emoji. The agent saves these details in an identity template, solidifying OpenClaw's status as one of the weirdest agentic platforms. It even has a soul template that instructs the agent that “You're not a chatbot. You're becoming someone.” This prompt structure makes the underlying model generate more customizable, conversational text than the average chatbot, which led to the creation of viral websites like Moltbook, a purported social network for AI agents, earlier in 2026.

An OpenClaw agent setting itself up on Windows
I let the agent choose its own name and personality (Credit: Microsoft/PCMag)

When I asked the agent to show off what it could do, it tried to run some commands but failed because of the sandbox. The commands in question were safe enough, but the default sandbox configurations err on the side of restricting the agent to be safe. It also took a screenshot of my desktop, commented on the weather forecast visible on my taskbar, and attempted to generate an image for this very article. (I didn't use any AI-generated images or text in this article!)

Later, the agent insisted it could use my web browser, so I asked it to open a browser tab and perform a web search. The browser didn’t appear. The agent spent several minutes running commands, reading help files, and repeatedly restarting the underlying OpenClaw gateway as it silently tried to troubleshoot why the browser tool wasn't functional. Eventually, it launched a web canvas view and navigated to a web page. The agent informed me that the browser tool wasn’t set up properly and that I could troubleshoot my OpenClaw configuration to get it working. I was both impressed that the agent managed to open a web page and slightly horrified that the AI had to spend several minutes examining the OpenClaw software to accomplish such a simple task.

You can use the Cron option to run your agent on a schedule. I asked the agent to delete all .exe files in my Downloads folder, for example, which could include an instruction to do this every day. But even with this simple task, the agent initially insisted that it couldn’t access my Downloads folder. I had to instruct the agent to use system.run tools to do so; the sandbox preferences pane in the application gave me that clue. The agent was then able to discover the right tool for the job. Still, I had concerns that the agent would make a mistake and delete the wrong file. It just isn't predictable, so I wouldn’t trust it with such a task.

Perhaps the most surprising takeaway from my testing is that the OpenClaw agents often don't seem to understand how OpenClaw itself works. Having to provide instructions or wait for it to troubleshoot itself seems to defeat the purpose. Apple Intelligence doesn't have this issue in macOS Golden Gate, which really crystallizes the experimental development state of OpenClaw.

The OpenClaw companion app on Windows showing a command blocked
My agent tried to be a little too helpful (Credit: Microsoft/PCMag)

Mobile Chat: Texting My Computer Worked, But the Agent Argued Back

I didn't have any issues sending messages to my OpenClaw AI agent via a Telegram bot, which is one of the recommended chat channels, alongside Discord, Google Chat, iMessage, and SMS. Assuming you leave your Windows PC running, you can text your agent from your phone and have it interact with your computer. At Microsoft Build, Nvidia CEO Jensen Huang appeared remotely to talk about how excited he was about the ability to text his PC using such tools.

From my phone, I asked the agent to look at my webcam and tell me if anyone was sitting in front of my PC. The agent insisted that it didn’t have access to my webcam and argued that this would violate privacy, so it would definitely not look at my webcam even if it could. The OpenClaw companion app did show the agent had webcam access, so this is yet another problem I could have attempted to troubleshoot.

I like the idea of having a powerful AI agent with a personality accessible via a chat app I already use, but the seemingly endless configuration process ruins the experience. Meanwhile, Apple Intelligence is a core part of Apple's OSes and works seamlessly across devices. That's the vision Microsoft should aim for.

An OpenClaw agent in Telegram
OpenClaw lets you hook up your agent to a variety of chat apps, including Telegram (Credit: Microsoft/Telegram/PCMag)

The Future of Windows Might Be Agentic, But OpenClaw Isn't Ready for Prime Time

As I mentioned at the top, the OpenClaw app just isn’t ready for regular Windows users. As of now, it's little more than a janky, open-source agent system that's just weird enough to be memorable. It's an early and experimental expression of Microsoft's desire to position Windows as a platform for AI tools and gives the company a way to test its sandbox, but the underlying technology still isn't practical.

My colleague Ruben Circelli called OpenClaw overhyped earlier this year, and his criticisms remain largely valid. The setup process remains complex despite the traditional app interface; token costs continue to rise; and it’s unclear what you should actually use OpenClaw for. The MXC sandbox represents a significant step forward for running the agents securely, however.

Microsoft clearly doesn't intend to deviate from OpenClaw any time soon. The company separately launched the more polished Microsoft Scout, a business-focused AI agent system that uses OpenClaw under the hood, and no longer seems to care about Copilot+ PCs. Whether it can transform its various OpenClaw projects into a compelling AI experience for all users remains a big question, but, for now, Apple is doing a far better job of selling agentic AI tools to more people.

About Our Expert