Claude Computer Use - why the Chrome plugin misses the point

The wrong tool for a very big problem

The tech world got handed a full symphony orchestra and decided they only need the second violin. That’s Claude’s Computer Use right now.

Anthropic built something that can see and interact with your entire digital workspace. Every application, every window, every pixel on your screen. Claude Sonnet 4.5 scores 61.4% on OSWorld, a benchmark measuring computer control capabilities across real applications. And what’s the first thing everyone builds? Chrome extensions. Browser plugins. Even Anthropic’s own Chrome extension limits Claude to just the browser.

We’ve apparently learned nothing from decades of digital fragmentation. The same fragmentation that undermines AI readiness across enterprises.

What a real workday actually looks like

Watch any knowledge worker for a day and the pattern is depressingly consistent.

The average employee has dozens of applications open. They switch between them constantly. Each switch breaks concentration. Each context change burns mental energy that doesn’t come back.

Your own workday: email, Slack, Excel, your CRM, a project management tool, the documentation wiki, your code editor, a banking portal, an analytics dashboard. You’re not working. You’re running a frantic relay race between tools that don’t talk to each other.

A browser extension is going to fix this?

That’s like putting a bandaid on a severed artery and calling it surgery.

What most people miss is this: Claude’s Computer Use doesn’t just automate clicking. It understands the visual language of software itself. Every application you use, whether that’s Slack, Excel, Salesforce, your IDE, or that proprietary tool your company built in 2003, they all speak the same visual language. Buttons look like buttons. Text fields look like text fields. Menus behave like menus.

Claude can read this language across every application at once. The Model Context Protocol (MCP) standardizes how Claude integrates with external tools and data sources, and adoption has exploded from around 1,000 servers to tens of thousands in a matter of months.

Not a browser automation tool. A universal software translator.

When I was growing up in Kenya, I watched my music teacher struggle with something called a “symphony desk.” A massive piece of furniture designed to hold all the sheet music for different instruments. Violin parts here, brass there, percussion in another drawer. The conductor had to physically shuffle between sections, losing the flow of the entire piece.

That’s our desktops now. Each application is a different section of the orchestra, physically separated, requiring constant shuffling. Computer Use changes this. One view. All instruments. Real conducting.

Why everyone built browser tools first

I think I understand why this is happening, even if I find it genuinely frustrating.

Fear of scope is real. A browser is contained, predictable, safe. You can’t accidentally delete system files or expose sensitive data outside the browser sandbox. It’s the kiddie pool of automation.

Chrome’s extension API is also mature. Well-documented. Thousands of examples. Why think harder when you can think easier.

And then there’s venture capital theater. “We’re building the Chrome extension for Claude” fits on a slide. “We’re rebuilding how computers interface with human intention” doesn’t. Guess which one gets funded faster.

The deeper problem is that we’re measuring the wrong things. We’re tracking “time saved on browser tasks” when we should be tracking “cognitive load eliminated from workflow fragmentation.” This exact disconnect is what drives the process failures we see in AI incidents.

The productivity stack Chrome extensions can’t reach

Your work data isn’t in your browser. It’s scattered across what I think of as the seven-layer productivity stack:

1. Communication layer: Slack, Teams, Discord, email
2. Documentation layer: Notion, Confluence, Google Docs, wikis
3. Data layer: Excel, Sheets, Airtable, databases
4. Development layer: VS Code, GitHub, terminal, Docker
5. Customer layer: CRM, support tickets, user analytics
6. Financial layer: QuickBooks, Stripe, banking portals
7. Proprietary layer: That custom tool only your company uses

A Chrome extension touches maybe one and a half of those layers. Computer Use coordinates all seven.

Think about what this actually unlocks.

The Monday morning ritual: instead of spending 45 minutes gathering data from six different tools for your weekly report, you describe what you need. The AI pulls from everywhere, assembles it, and presents it for your review.

The customer fire drill: a support ticket comes in. Instead of jumping between the CRM, codebase, logs, and documentation yourself, the AI instantly correlates the issue across all systems and gives you the full picture.

The proposal process: no more copy-pasting between pricing spreadsheets, document templates, and CRM data. One request, full coordination, complete proposal.

This isn’t about saving minutes. It’s about preserving cognitive flow.

What’s actually coming

Most companies aren’t ready for this. Not technically. Technically is the easy part. Culturally. As I’ve written when exploring how to communicate AI changes effectively, the human side is always harder than the technical side.

We’ve spent decades building walls between applications. Security walls. Process walls. Departmental walls. Computer Use makes those walls visible in ways that probably terrify the people who built their careers managing them.

Building Tallyfy showed me something about this. The companies that break down these walls first don’t just get more efficient. They develop fundamentally different capabilities. They start solving problems that siloed companies can’t even see.

Update: Anthropic released Claude Cowork, which is exactly this vision. A general-purpose agent that coordinates work across your entire digital workspace, not just browsers. It can read, edit, and create files while planning multi-step tasks that run for 30 minutes or longer. The shift from browser automation to workspace orchestration is happening faster than I expected.

Why should you need to know which application contains which data? Why should you care whether something lives in Slack or email or Notion? You want answers, not treasure hunts.

The winners will be the ones who realize Computer Use isn’t about automating browsers. It’s about making the entire concept of application switching obsolete.

In five years, we’ll look back at Chrome extensions for AI the way we look at WAP browsers for mobile. A necessary stepping stone that completely missed the point.

The real shift isn’t in making browsers smarter. It’s in making the computer itself readable by AI. When that happens, when AI can truly see and coordinate everything we do digitally, the idea of manual application switching will seem as antiquated as hand-copying manuscripts.

But sure. Let’s build another Chrome extension.