Week 3

Chrome Use, Computer Use & MCPs

NYU · Whatever Generative AI Is Doing Now · 2026

SCHEDULE

Week 1 Install and Claude Basics
Week 2 Research and Writing with Claude
Week 3 Chrome Use, Computer Use & MCPs
Week 4 Build your own software part 1
Week 5 Build your own software part 2
Week 6 Build your own software part 3
Week 7 Demo Day

TODAY

1. Browser automation with Chrome Use
2. Desktop control with Computer Use
3. MCPs — how all of this actually works under the hood
4. Hands-on exercises

Section 1

Browser Automation with Chrome

Claude can’t always access every URL from within Claude Code. Sometimes you need to open a webpage.

CHROME USE

Chrome Use lets Claude interact with your Chrome browser.

Chrome USE

It can navigate to URLs, read what’s on a page, click buttons, fill out forms, take screenshots, even record GIFs of what it’s doing.

SETTING IT UP

1. Install the Claude in Chrome extension from the Chrome Web Store
2. From Warp, run `claude --chrome`. The first step will require you to accept permissions.
3. Quit CC (ctrl + c twice)
4. Now run `claude`
5. From CC, run `/chrome`. Make sure Status is Enabled, And Enabled by Default: yes

HOW IT WORKS

Chrome Use reads the page’s DOM — the HTML source, not just what you see on screen.

WHAT I ACTUALLY USE IT FOR

Testing: check if something I built actually works in the browser
Form filling: those repetitive web forms that make you want to close your laptop
Screenshots: capture the state of pages for documentation
Web scraping: extract structured data from pages that don’t have APIs

Exercise

Automate a Browser Task

EXERCISE

Use Claude with Chrome Use to:

• Navigate to a site, extract key information, and summarize it
• Fill out a form using data Claude already knows about you
• Take screenshots of multiple pages and compare them

You can also ask Claude to record a GIF of the whole process

TIPS

• Start simple: “use /chrome; go to this URL and tell me what you see”
• It sometimes struggles with really complex single-page apps
• If something fails, try having Claude take a screenshot first — it helps it understand the layout
• Don’t trigger alert dialogs!
• Be patient with it. The first attempt doesn’t always work perfectly but it usually figures it out.

Section 2

Computer Use

Like /chrome, but for your entire screen.

WHAT IS COMPUTER USE?

Computer Use is the next step up from Chrome Use. Instead of your browser, Claude can control your whole desktop — seeing the screen, moving the cursor, clicking, typing. Like a human sitting at your computer.

WHAT IS COMPUTER USE?

The big difference is it works with any application. Photoshop, Finder, system settings, whatever. It takes screenshots, figures out what to do, and sends mouse and keyboard inputs.

HOW IT WORKS

1. Claude takes a screenshot of your screen
2. It looks at the screenshot and identifies what’s there
3. It decides what to click or type
4. It sends that input
5. It takes another screenshot to see what happened
6. Repeat

HOW IT WORKS

It’s basically a loop: see → think → act → see. Which honestly sounds simple but is remarkable when you watch it happen.

WHAT’S THE DIFFERENCE?

Chrome Use:
• Browser only
• Reads the actual page structure (fast, precise)
• Works through a Chrome extension

WHAT’S THE DIFFERENCE?

Computer Use:
• Any application on your screen
• Uses screenshots (slower, more visual, less precise)
• Works through screen capture + input simulation

USE CASES

Creative tools: have Claude interact with Photoshop, Blender, After Effects, things like that
Multi-app workflows: copy data between apps that don’t talk to each other
Legacy software: automate tools that have no command line or API
Testing: verify that any UI works the way it should
Demonstrations: show someone how to use software, step by step

Basically anything where you’d normally have to click through a bunch of menus yourself.

LIMITATIONS

• It’s slower than Chrome Use or direct APIs — every action requires a screenshot round trip
• It can misread UI elements, especially small text or overlapping buttons
• It needs clear visual feedback — if the screen doesn’t change, it doesn’t know the action worked
• It burns through tokens fast because of all those screenshots
• Sometimes it gets stuck in a loop clicking the same wrong spot over and over

LIMITATIONS

It can be quite frustrating, but this technology will only get better.

Exercise

Try Computer Use

EXERCISE

Use Computer Use to interact with a desktop application — something that isn’t a browser.

Ideas:
• Open a design tool and have Claude describe what it sees
• Navigate through system preferences to find and change a setting
• Have Claude take a screenshot of your desktop and describe your setup
• Try automating something that spans multiple applications

Pay attention to where it works and where it falls apart. That’s the interesting part.

TIPS

• Close windows you don’t need
• Make sure whatever app you want to control is actually visible, not minimized
• If it keeps clicking the wrong spot, you can give it more specific instructions
• Be patient. Like, really patient.
• If you see it looping on the same failed action, just stop it and try a different approach.

AND IT’S NOT JUST MCPs

So this isn’t an MCP, but I think it shows the same idea. I built a project called adobe-hut — it’s a collection of Claude skills and agents specifically for Adobe Creative Suite.

After Effects scripting, Photoshop automation, Premiere, InDesign, Illustrator. The skills teach Claude how ExtendScript works, what the APIs look like, what the gotchas are. So instead of Claude guessing at After Effects code, it actually knows the correct syntax.

The point is: you can use skills to make Claude an expert in tools you already use. You don’t have to switch to new tools — you just make your existing ones smarter.

Section 3

Model Context Protocol

So here’s the thing — Chrome Use and Computer Use are both MCPs.

WHAT IS MCP?

MCP stands for Model Context Protocol. A universal way to plug new capabilities into Claude.

Without MCP, Claude can really only read and write files. But with MCP, it can manage databases, control your browser, talk to APIs, generate images. Everything you just used today? That was MCP. It’s all the same underlying system.

HOW IT WORKS

You → ask Claude to do something
Claude → recognizes it needs a specific tool
MCP Server → provides that tool
External Service → does the actual work

Claude doesn’t directly call APIs or control your browser. It talks to these MCP servers, which act as translators between Claude and everything else.

MCPs I ACTUALLY USE

Notion MCP: reads and writes your Notion workspace. If you live in Notion for project management or mood boards, this is a game changer
Fal.ai: image generation directly from Claude Code. You can say “generate a reference image of X” without leaving the terminal
Context7: pulls live documentation for any framework or library. No API key needed, super easy to install

MCPs FOR CREATIVE TOOLS

Blender: 18,000+ stars on GitHub. Natural language 3D scene creation. This one’s insane.
DaVinci Resolve: AI-assisted video editing. Full timeline and project control.
TouchDesigner: control TD projects from Claude. Real-time visuals people, this one’s for you.
Max/MSP: Claude can understand and generate Max patches
Unreal / Unity: game engine MCPs are surprisingly mature

INSTALLING AN MCP

/install-mcp

You can also add them manually to your settings files:

~/.claude/settings.json (global — available everywhere)
.claude/settings.json (per-project — only in this folder)

Most MCPs are npm packages or Python scripts.

WHAT THE CONFIG LOOKS LIKE

A typical MCP entry looks like this:

"mcpServers": {
  "server-name": {
    "command": "npx",
    "args": ["-y", "package-name"],
    "env": { "API_KEY": "..." }
  }
}

You don’t need to write code to use MCPs. You just install them.

WHERE TO FIND MORE

There are thousands of MCPs now:

github.com/modelcontextprotocol/servers — the official reference servers
mcp.so — community directory, the main hub
glama.ai/mcp/servers — curated list with reviews

Think about the tools you already use. Google Drive, Figma, Slack, GitHub — there’s probably an MCP for it.

CHOOSING THE RIGHT TOOL

1. MCP / API — fastest, most reliable, direct data access. If one exists, use it. (Databases)
2. Chrome Use — great for web apps, reads the DOM, pretty precise (Web apps)
3. Computer Use — works with anything on screen, but slower and more fragile (Mac apps)

Computer Use is the “break glass in case of emergency” option.

Exercise

Install Your First MCP

EXERCISE

Pick one or two of these and install them right now:

Notion MCP — if you use Notion, this one’s a no-brainer. You’ll need a Notion API token.
Fal.ai — generate images from Claude Code. You’ll need a fal.ai account and API key.
Context7 — documentation lookup. No API key, easiest one to install if you want a quick win.

Ask Claude for help, and use cases.

Automate the friction between you and the work you actually want to do.

Homework

Due Next Week

HOMEWORK 1: BUILD YOUR MCP TOOLKIT

Install at least 3 MCP servers that are actually relevant to your practice.

For each one, write down:
• What it does
• Why you picked it
• One real task you completed with it
• What worked and what didn’t

HOMEWORK 2: AUTOMATE A REAL WORKFLOW

Find a repetitive task in your creative practice — something you do all the time that makes you want to scream — and automate it. Use any combination of MCPs, Chrome Use, or Computer Use.

HOMEWORK 2: AUTOMATE A REAL WORKFLOW

Examples:
• Gather reference images from multiple sites into a folder
• Update your portfolio site content from a Notion database
• Export and organize project files across applications

Document what you did.