Claude Code and Chrome: testing your website pages in a real browser

Claude Code can drive a real web browser through the Claude in Chrome extension. Once connected, Claude can open a page on your local copy or staging site, click through it, fill in a form, read the browser console for errors and report what it found, then go back to the code and fix the cause. For website work, that closes the gap between "the code looks right" and "the page works."

The trade-off is trust. Anthropic's docs say Claude "shares your browser's login state, so it can access any site you're already signed into." That makes it useful for testing logged-in areas and risky anywhere you are signed in with real authority, such as your live CMS, ad accounts or payment dashboards. This guide covers setup, the website checks it does well, and the permission settings that keep it pointed at the right pages, as of October 2026.

What the integration is

Claude in Chrome is Anthropic's browser extension. On its own it lets Claude act in your browser from a side panel. Claude Code "integrates with the Claude in Chrome browser extension to give you browser automation capabilities from the CLI or the VS Code extension." Anthropic's glossary calls Chrome an integration that connects to a surface, not a surface of its own. The capabilities Anthropic lists include:

• Live debugging: "read console errors and DOM state directly, then fix the code that caused them"
• Web app testing: "test form validation, check for visual regressions, or verify user flows"
• Design verification against a mockup
• Data extraction from pages into a local file
• Task automation such as data entry and form filling
• File uploads and session recording as GIFs

The desktop app has a separate built-in Browser pane that uses a clean profile with no saved logins. If you only need to preview pages you are building, that is the lower-risk option, covered in our guide to the Claude Code desktop app.

Requirements and browser support

Which browsers. Anthropic's pages disagree here. The Claude Code Chrome documentation says the integration "works with Google Chrome and Microsoft Edge," and that Claude Code also sets up the connection in Brave, Arc, Vivaldi and Opera. Anthropic's August 2026 announcement that the extension is generally available says it "doesn't run on other Chromium browsers or on mobile yet." Google Chrome is the one browser both pages support, since Edge is itself a Chromium browser. If you rely on Edge or another Chromium browser, test it and check Anthropic's current pages.

Anthropic describes Claude in Chrome as generally available. The two first-party sources give different timing, so check Anthropic's pages if the date matters to you.

Setting it up from the terminal or VS Code

1. Install the extension from the Chrome Web Store and sign in to the same Claude organization you use for Claude Code. If Claude Code needs the browser and can't find the extension, it offers an install prompt.
2. Start a session with Chrome. In the terminal, run claude --chrome. In VS Code, type @browser followed by the task.
3. Check the connection. Run /chrome. The integration is working when it shows "Status: Enabled" and "Extension: Installed." The same menu manages permissions, reconnects the extension, and picks a browser when more than one is connected.
4. Decide on "Enabled by default." You can turn Chrome on for every session from /chrome. Anthropic notes this "increases context usage since the browser tools and their instructions are always loaded." Many teams leave it off and connect only when a task needs a browser.
5. Run a harmless first task, such as opening your local site's homepage and listing any console errors.

A repository can't switch Chrome on for you. Anthropic blocks a project's own settings file from enabling it, because "a repository you check out must not be able to connect Claude to your browser."

Website checks Claude can run in the browser

These are the jobs where a browser adds real value. Run them on a local copy or a staging site.

Two habits make these checks reliable. Tell Claude what to look for rather than asking for "all console output," which Anthropic notes can be verbose. And ask for the evidence, such as the exact error text or the URL where a step failed, so you can confirm it yourself. For the full pre-deploy list, see our website QA routine, and for forms specifically, our guide to building and fixing contact forms.

Browser checks complement, not replace, crawl-based SEO checks. Our Claude Code for technical SEO guide covers the crawler side.

Permissions: what Claude may do on which sites

"Site-level permissions are inherited from the Chrome extension." When Claude Code wants to act, the prompt starts with "Claude in Chrome wants to" and can allow actions on that site for the session. Anthropic's permissions guide lists three choices: "Allow this action", "Always allow actions on this site" and "Decline". Even with Always allow, Claude still asks before downloading a file, entering potentially sensitive information or granting authorizations.

Some extra rules apply:

• Adult content websites and known pirated content sites are blocked by default, and Claude asks permission before accessing financial sites.
• In auto mode, when the auto mode classifier approves a browser call, the extension skips its own per-site check for that call, unless your permission rules deny the site to Claude in Chrome.
• In plan mode, a permission prompt appears before Claude records a GIF, opens a new tab or runs a shortcut.
• Team and Enterprise admins can set allowlists that restrict Claude to approved sites, and blocklists.

For a website team, a sensible default is to allow your local development address and your staging domain, and nothing else, until a task needs more.

The risks, and how to reduce them

Your logins are Claude's logins. If you are signed in to the live CMS as an administrator, Claude can act there too. Consider a separate Chrome profile for Claude work, signed in only to staging, with an account that has the lowest role the task needs. Our website security checklist covers account roles more broadly.

Prompt injection. Web pages can contain hidden instructions aimed at AI agents. Anthropic says Claude in Chrome has safety classifiers: one "checks incoming content for injection attempts, and another checks every action Claude takes before it runs." It also says plainly that "The risk is not zero." Anthropic's own testing, published with the GA announcement, reports that attacks reaching the model succeeded 3.8% of the time against Opus 5 before additional safeguards. With probes and the classifier in place, it reports no successful attacks against Sonnet 5, Opus 5 or Mythos 5, and a 0.3% rate against Fable 5. Keep Claude on sites you control. Our guide to prompt injection risks on websites goes deeper.

Real submissions. A form submitted on your live site creates a real lead, email or order. Test forms on staging, or tell Claude explicitly not to submit on production.

Recordings and uploads. A GIF "captures everything visible in the browser, including account details on logged-in pages." Review it before sharing outside your team. File uploads are capped at 10 MB in total per upload, and Claude refuses credential-like files such as .env and key files.

The live check that lies. When we rebuilt eseospace.com, our CDN once served a stale copy of a page after a deploy, making a finished change look missing. A browser check of a live URL shows what the cache serves, which may not be what you deployed. Check the build output and git history before trusting a live check.

For a broader list of controls, see our guide to Claude Code on production websites.

Troubleshooting the common problems

• "Browser extension is not connected": the extension may be signed in to a different claude.ai account or organization than Claude Code. Run /status to see Claude Code's organization.
• Commands stop working: a JavaScript alert or confirm box blocks browser events. Dismiss it, then tell Claude to continue.
• It worked, then stopped: the extension can go idle in long sessions. Run /chrome and choose "Reconnect extension."
• Not detected the first time: restart Chrome so it picks up the connection settings Claude Code installed.

Conclusion

Connecting Claude Code to Chrome lets it test what it builds: open the page, use the form, read the errors and fix the cause. That is valuable for website QA, as long as Claude works on staging, in a browser profile with limited logins, with site permissions narrowed to what the task needs and a person confirming the results. Keep it away from live admin sessions, payment pages and anything you can't easily undo. If you want browser testing built into your site's review process, our Claude Code development team can set it up. More guides are in our Claude Code Guide.

Frequently asked questions