My AI minions keep getting more capable, and honestly, it’s both exciting and slightly terrifying. One step closer to world domination… but first, too many meetings.
Claude in Chrome is Anthropic’s browser automation extension, now available in beta to all paid Claude subscribers. I’ve been testing it over the past week. The short version: it works. The longer version involves me being extremely careful about what I actually let it do.
How It Works
The extension operates as a side panel in your browser. Claude can read, click, and navigate websites alongside you, seeing what you see and taking actions when you ask. You control access through site-level permissions, granting or revoking Claude’s ability to interact with specific websites anytime.
A few features worth noting. The “Ask before acting” mode has Claude create a plan for your approval before executing. You can record workflows by performing steps yourself, and Claude learns to repeat them. This is useful for repetitive browser tasks that follow the same pattern. You can even schedule recurring tasks to run automatically on a daily, weekly, or monthly basis.
It supports Haiku 4.5, Sonnet 4.5, and Opus 4.5, so you can choose your model based on the complexity of what you’re asking it to do. I’ve been running most tasks on Haiku since it’s faster for simple clicking tasks.
For full details on features and safety considerations, check out Anthropic’s official guide: Getting Started with Claude in Chrome
Getting Started
You’ll need a paid Claude subscription (Pro, Max, Team, or Enterprise). From there:
- Visit the Chrome Web Store and search for “Claude in Chrome”
- Click “Add to Chrome” to install
- Sign in with your Claude account credentials
- Pin the extension by clicking the puzzle piece icon in your toolbar, then the pin next to Claude
- Click the Claude icon to open the side panel and start giving it tasks
Using Edge Instead of Chrome
Despite the name, Claude in Chrome works perfectly in Microsoft Edge since Edge is Chromium-based. To install:
- Open Edge and go to the Chrome Web Store
- If prompted, click “Allow extensions from other stores” in the banner at the top
- Search for “Claude in Chrome” and install it the same way you would in Chrome
Everything works identically once installed.
My Current Comfort Level
Here’s where I’m at with personal, low-stakes tasks only. Gmail cleanup. Simple admin work. And my new favorite: generating Plaud meeting notes so I don’t have to click through each recording manually.
That Plaud task is a perfect example of where browser automation shines. I pay for Plaud’s unlimited plan, but it doesn’t auto-generate transcriptions. So I end up with a list of recordings showing timestamps instead of titles, and I have to click into each one, hit generate, wait, repeat. It’s not hard. It’s just tedious enough that I put it off.
I pointed Claude at my Plaud account with simple instructions: go through the last few weeks, find any meeting over 15 minutes that hasn’t been generated yet, and generate the notes. I gave it one tip. Ungenerated meetings show as date/time stamps instead of actual titles. 55 steps later, it was done.
Here’s why I’m comfortable with this particular task. Claude leaves the page before the notes are actually generated. It’s clicking the generate button, not reading the content. The actual transcription happens after it moves on. That distinction matters to me.
What I’m not doing yet is anything work-related. No client data. No sensitive systems. Nothing that matters. I’m still building trust with this tool, and that means testing it on things where a mistake costs me nothing but mild annoyance.
The Security Reality
Anthropic is refreshingly transparent about the risks. The biggest concern is prompt injection, which involves hidden instructions on websites that could trick Claude into doing something you didn’t ask for. Before safety measures, their internal red-teaming found attacks succeeded 23.6% of the time. With current defenses, that’s down to 11.2%. Better, but not zero.
The extension asks for confirmation before high-risk actions like publishing, purchasing, or sharing personal data. That’s reassuring, but I’m still treating this as a tool that requires supervision on anything that matters.
The Downstream Workflow
The downstream automation is where this gets interesting. I have a Zapier workflow that monitors for new Plaud transcriptions and routes them to the appropriate client folders in OneDrive. Then when I’m prepping for a client task, Copilot can surface relevant meeting notes without me digging through folders. The combination of Claude handling the tedious generation clicks, Zapier handling the filing, and Copilot handling the retrieval means my meeting notes actually become useful instead of sitting in a pile I never look at.
What’s Next
Next test on my list: ordering my standard Wednesday Mendocino Farms lunch. Low stakes, repeatable task, good way to see how it handles a checkout flow. Maybe. I’m still deciding if I trust it with my sandwich order. I’m a picky eater. What if it forgets my pickles? The horror!
For IT leaders watching these tools evolve, my advice is simple. Test on your own stuff first. Pick tasks that are boring, repetitive, and completely fine if they fail. Build your understanding of what the tool actually does before you point it at anything that matters.
The AI minions are getting more capable every month. Learning to work with them safely is becoming a core skill.