Claude Code's Built-In Browser: What It Is and How to Use It

On July 10, 2026, Anthropic shipped something Claude Code users had been asking for since the tool first launched: a browser that lives inside the app itself. Instead of tabbing out to Chrome to check a docs page, compare a design mock, or verify that a UI change actually rendered correctly, Claude Code on desktop can now open, navigate, and interact with any website in a sandboxed window — without ever leaving your terminal-and-editor workflow.

It shipped as part of Claude Code's Week 28 release (versions 2.1.202 through 2.1.206), alongside a rebuilt /doctor command and tighter auto-mode safety guardrails. If you're using Claude Code daily — for research, debugging, or agentic dev workflows — this is one of the more consequential UX changes of the month. Here's what it actually does, how it differs from the existing Claude for Chrome extension, and how to start using it today.

What the Built-In Browser Actually Does

The built-in browser gives Claude the same kind of visual and interactive access to any website that it's had to your local dev server preview since Claude Code added live app previews. Concretely, Claude can now:

Open documentation for a library or API mid-task, read the actual page content, and apply what it finds without you copy-pasting URLs into the chat
Compare a Figma or design mock against your rendered UI, side by side, and flag mismatches
Click through multi-step flows on a real website — filling forms, following links, testing signup or checkout paths — the same way it clicks through your local preview
Read rendered JavaScript-heavy pages that a simple fetch-and-parse tool would choke on, because it's a real browser context, not a scraper

You launch it with Cmd+Shift+B on macOS or Ctrl+Shift+B on Windows, or from the Views menu in the desktop app. It opens as its own pane alongside your terminal and editor, so you can watch Claude navigate a page in real time — the same transparency you get when it drives your local dev preview.

This closes a real gap. Before this release, if you wanted Claude Code to check a third-party site — pull the latest params from a payment provider's API docs, verify how a competitor's landing page structures its pricing table, confirm a webhook payload example matches what you're sending — you either fed it a WebFetch-style tool with no interactivity, or you did it yourself and pasted the results back in. Now it can just go look.

How It's Different From the Claude for Chrome Extension

Anthropic already shipped a Claude for Chrome extension that lets Claude act inside your actual browser, using your logged-in sessions to work across Gmail, Notion, Google Docs, and other authenticated apps. It's easy to assume the new built-in browser is redundant with that — it isn't, and the distinction matters for how you use each one.

Claude Code built-in browser	Claude for Chrome extension
Where it runs	Inside Claude Code desktop, sandboxed	Inside your actual Chrome browser
Session state	Stateless by default (configurable to persist)	Shares your logged-in Chrome sessions
Best for	Docs lookup, design comparison, testing public flows, one-off research mid-task	Automating tasks across apps you're already logged into (email, docs, project tools)
Isolation	Sandboxed from your personal browsing data	Runs with your real browser profile and cookies
Typical trigger	Automatic, when Claude decides it needs to check something during a coding task	Explicit, when you ask Claude to do something in your actual accounts

The short version: use the built-in browser when you want Claude to look something up or test something without touching your real accounts. Use the Chrome extension when the task genuinely requires being logged in as you. If you're building a feature and want Claude to cross-reference the Stripe API docs while it works, that's the built-in browser. If you want Claude to pull data out of your Gmail inbox, that's the Chrome extension.

Safety and Permissions: What You Should Know Before Using It

Anthropic built this with the same "ask before acting" philosophy that governs Claude Code's file edits and shell commands. A few things worth understanding before you turn it loose:

Per-site permission prompts. The first time Claude attempts to act on a specific website — clicking a button, filling a form, submitting anything — you get a prompt: allow once, always allow for this site, or deny. This mirrors how Claude Code already asks before running a shell command or editing a file outside your expected scope, so the mental model should feel familiar if you've used Claude Code's permission system before. Sandboxed and stateless by default. Unlike the Chrome extension, the built-in browser doesn't inherit your personal login state. Sessions don't persist between uses unless you explicitly configure that — which matters if you're testing a signup flow and don't want stale cookies affecting the result. Safety classifiers review actions on external sites. Beyond the allow/deny prompts, Anthropic runs safety classifiers over the actions Claude takes in the browser, adding a layer of review specifically because this feature gives Claude reach beyond your local filesystem and dev server — the traditional trust boundary for Claude Code.

If you manage Claude Code at a team level, it's worth checking your settings.json permission configuration to see whether you want browser actions gated the same way as other tool calls, or whether you want to restrict which domains it can reach at all.

How to Turn It On and Start Using It

Getting started takes about thirty seconds:

Update Claude Code to a version in the 2.1.202+ range (or later — check with claude --version, or just let auto-update handle it).

Open the desktop app — this feature is desktop-only for now, not available in the terminal-only CLI.

Launch the browser with Cmd+Shift+B (Mac) or Ctrl+Shift+B (Windows), or via View → Browser in the menu.

Just work normally. You don't have to manually drive it — during a coding session, if Claude decides it needs to check documentation, compare a rendered page, or verify a flow, it'll open the browser pane itself and you'll see it navigate live.

Approve site access the first time it hits a new domain. Choose "always allow" for sites you'll want Claude checking repeatedly (your own docs site, a frequently-referenced API reference), and "once" for anything ad hoc.

If you don't see the option, confirm you're on the desktop app (not the CLI-only install) and running Ubuntu 22.04+/Debian 12+ if you're on Linux — desktop app support for those distros also shipped this cycle, in beta.

What This Signals About Where Claude Code Is Headed

This release didn't land in isolation. The same Week 28 update rebuilt /doctor from a read-only diagnostic into a command that actually fixes what it finds — unused skills, misconfigured MCP servers, slow hooks — and tightened auto-mode so it now blocks tampering with session transcripts and pauses before running an rm -rf on a variable it can't resolve from context.

Read together, the pattern is consistent: Anthropic is expanding what Claude Code can do autonomously (browse the live web, fix its own configuration) while simultaneously tightening the guardrails around consequential actions (site permissions, transcript protection, destructive-command checks). That's the trajectory to watch if you're evaluating Claude Code for team-wide rollout — more capability is shipping alongside more scoped safety, not instead of it.

For teams preparing for the Claude Certified Architect (CCA) exam or building deeper Claude Code fluency, understanding how these agentic capabilities are scoped and permissioned isn't just trivia — it's increasingly what separates surface-level Claude Code users from people who can architect safe, production-grade agent workflows.

Key Takeaways

Claude Code desktop now has a sandboxed, built-in browser (Cmd/Ctrl+Shift+B) for reading docs, comparing designs, and testing web flows without leaving the app
It's stateless and sandboxed by default — distinct from the Claude for Chrome extension, which uses your real, logged-in browser sessions
Every new site requires an explicit allow/deny decision the first time Claude acts on it, reviewed by safety classifiers
It shipped alongside a smarter /doctor and stricter auto-mode safety checks — part of a broader pattern of more autonomy paired with tighter guardrails
Desktop-only for now (2.1.202+), including new Linux beta support on Ubuntu 22.04+/Debian 12+

Next Steps

Want to go deeper on what else shipped this cycle, or get hands-on with Claude Code fundamentals before layering in agentic browser workflows? Check out our guide to getting started with Claude Code and how to build custom Claude Code skills — and if you're prepping for the Claude Certified Architect exam, AI for Anything's CCA practice tests are built around exactly these kinds of real-world Claude Code capability changes.

Sources:

Claude Code's Built-In Browser: What It Is and How to Use It (2026)