Claude in Chrome and website feedback: which your agent needs, and when

This week we shipped a new hero on the Skyelight homepage. Claude Code built it, opened it in a browser, took screenshots at three sizes, and everything looked right. Later that day I opened the live site and the animated dots behind the headline were huge and soft, like a photo zoomed too far.

Claude in Chrome, the extension Claude Code drives when you run claude --chrome, can open your site, click through it and read the console. None of that would have caught this. The page threw no error. Nothing was broken in a way a browser reports. It only looked wrong to someone who knew what it looked like the day before.

That's the split. Claude in Chrome is how your agent checks its own work. Website feedback is how the people who aren't in your terminal tell it what's wrong. You want both, and one doesn't stand in for the other.

What does Claude in Chrome do?

It's Anthropic's browser extension, and Claude Code connects to it to use your browser. Once it's on, Claude opens tabs in a visible Chrome window, signed in to everything you're signed in to, and can:

• Read console errors and the page's structure, then fix the code behind them
• Fill in forms and click through a flow to test it
• Check that a build matches its Figma mock
• Take screenshots and record a flow as a GIF
• Pull data out of pages, and upload files from your machine

You need a direct Anthropic plan (Pro, Max, Ultra, Team or Enterprise), Chrome or Edge, and version 1.0.36 or later of the extension. Start Claude Code with claude --chrome, or run /chrome and choose Enabled by default. Anthropic notes that leaving it on by default loads the browser tools into every session, which uses more context.

What can't it see?

What the page was meant to look like. An agent compares the page with what you told it. Nobody had told it the dots were supposed to be small. When I did, Claude Code found the cause in minutes: moving the film into the hero had made the area behind the dots 900 pixels tall instead of 710, and the dots grow with it. It opened the old and new versions side by side, measured both and put the height back. Its own check that morning had missed it for a plainer reason. The browser it used couldn't draw the animation at all, so the screenshots showed a clean, empty background.

Anyone else's eyes. Claude in Chrome runs in your browser, on your machine, under your login. Your designer, your product manager and your client aren't in that session. Their notes still arrive the old way: a screenshot in Slack, or a long email that describes a page you have to find again.

Taste. “The headline feels a little long on mobile” isn't an error any tool can read from the page. It's a judgment, and it belongs to whoever is reviewing the work.

When to use Claude in Chrome

Use it whenever your agent needs to check its own work in a real browser:

• Testing a change on localhost before you commit it
• Chasing a console error or a failed request back to the code
• Walking a sign-up or checkout flow after a change
• Checking a fix someone asked for, before telling them it's done

It works best when the agent knows what “done” looks like, so put the expected result in your prompt.

When you need website feedback

You need it the moment someone other than you has an opinion about the page. That's every client review, every design pass and every “can you look at the staging site” message.

In Skyelight, the reviewer presses C, clicks the element and writes what's wrong. The comment stays pinned to that element as a thread, with the page, the browser and any errors from just before. Add Skyelight Build to your app and every thread also carries the source file and line behind the element. The thread has an owner and a status, so a review round of 20 comments doesn't turn into 20 copy-pastes.

How do they work together?

Give Claude Code both: the Skyelight MCP for the feedback, and Claude in Chrome for the browser. Then a review round runs like this:

1. A reviewer pins a comment on the page.
2. The agent reads the thread over MCP, with the element and the source line, and opens the right file.
3. It makes the change.
4. It opens the page with Claude in Chrome to check the fix, the console included.
5. It replies on the thread, and the reviewer sees the answer where they asked.

The person who noticed is the person who closes it. If you haven't connected an agent to Skyelight yet, here's the setup. It takes one command.

Frequently asked questions

An agent with a browser can tell you the page works. Only the people looking at it can tell you it's right.