Anthropic has integrated a full browser inside Claude Code. The agent can now open, read, click, and type on external web pages directly from the development environment, without leaving the editor. The browser works like a normal tabbed browser and opens with a keyboard shortcut.

What changes for daily Claude Code users. Before, you had to manually copy documentation, issue trackers, and external references, or switch windows every time. Now Claude opens and reads them on its own: it can search for an error on Stack Overflow, check a library’s documentation, or read the page of an API it’s integrating. As we reported on July 11, after security fixes Claude Code has moved toward autonomy: this browser is the next step.

Security controls are strict. Every write action on external sites goes through a classifier: Claude cannot buy anything, create accounts, or bypass CAPTCHAs without explicit consent. The browser runs on a clean profile with no saved logins. Organizations can limit access to an allowed sites list or disable the browser entirely. If Claude needs to act within a logged-in session, Anthropic recommends the dedicated Chrome extension.

In detail

The context

When Anthropic released auto mode as default on July 11, it signaled a clear direction: Claude Code becomes an agent that works independently for long sessions, not an assistant waiting for every command. The integrated browser is the logical continuation: an agent that writes code needs to read documentation, check issues, search for solutions. Before you did it and pasted it to him; now he does it himself.

It’s also an indirect response to OpenAI’s shutdown of Atlas, announced on July 10. While OpenAI decided to concentrate browser capabilities inside ChatGPT and the Chrome extension, Anthropic brings them inside the code editor itself. Two different strategies for the same problem: where does an agent that needs to read the web live.

How it works

The browser uses the same tools that Claude Code already uses for local app previews, but with an additional security layer. Every attempt to write to an external site (fill a form, click a submit button, modify content) is intercepted by a classifier. Read actions (open a page, scroll, read text) pass without filters.

The browser profile is isolated: no saved cookies, no authenticated session. If Claude needs to act within an account (for example, create a pull request on GitHub or modify a cloud resource), the user must use Claude’s Chrome extension, which runs in the normal browser with active sessions.

Enterprise organizations can configure a domain allowlist: Claude can only open those sites, and everything else is blocked upstream. Or they can disable the browser entirely via centralized policy.

What it means in practice

For developers using Claude Code, the immediate use case is documentation. Instead of manually searching for a library’s syntax, copying and pasting it into the prompt, ask Claude to open the docs page and read the section he needs. The same goes for issue trackers: Claude can open a bug report on GitHub, read the thread, and suggest a solution based on what others have tried.

The limitation is that Claude cannot buy, register, or authenticate. This cuts off many complete web automation scenarios (for example, ordering cloud resources, configuring external services, modifying published content). For those cases, the Chrome extension remains the right tool. The integrated browser is designed for read and navigation tasks, not transactional automation.

Broader implications

The choice to integrate the browser directly into the editor rather than a separate app is a signal about where Anthropic thinks development agents live: within the work context, not in a separate window. It’s consistent with the auto mode direction: fewer manual steps, more autonomy, but with granular controls to prevent damage.

The comparison with OpenAI is interesting: Atlas was a standalone product, and OpenAI chose to shut it down and consolidate everything in ChatGPT. Claude Code goes the opposite direction: every environment (coding, browsing, preview) becomes part of the same tool. Which approach scales better will be determined by real-world use in the coming months.