Anthropic adds sandboxed in-app browser to Claude Code desktop app

Anthropic has launched an in-app browser for Claude Code's desktop application, enabling the AI agent to interact with web content during development workflows.

According to the ClaudeDevs account on X, the browser allows Claude to "pull up docs, designs, or any other site" and "read, click through, and interact the same way it does with your local dev servers." The feature is sandboxed and configurable, with users controlling whether browsing sessions persist.

How it differs from the Chrome extension

Anthropic's documentation distinguishes the in-app browser from its existing Claude extension for Chrome. The in-app browser operates with a clean browser profile, separate from personal browsing data, with no saved logins or browsing history. This isolation makes it suitable for building and testing applications on sites that don't require user authentication.

For tasks requiring logged-in sessions, Anthropic recommends using the Claude extension in Chrome, which shares the browser's login state and can act as the user in authenticated contexts.

Context: AI-powered browsing competition

The announcement comes during a week of AI browser developments. OpenAI announced plans to discontinue its ChatGPT Atlas browser in favor of an in-app browser within the ChatGPT desktop app. Both Anthropic and OpenAI maintain Chrome extensions alongside their desktop browser integrations.

Multiple new AI-focused browsers have launched on Mac recently, reflecting growing interest in agentic web browsing capabilities that allow AI assistants to navigate and interact with web content autonomously.

What this means

The sandboxed browser architecture addresses a key challenge in AI agent development: giving models web access without compromising user security or privacy. By maintaining separation between the AI's browsing profile and personal browser data, Anthropic enables developers to let Claude interact with web resources during coding tasks while maintaining control over session persistence and authentication scope. This approach balances the utility of web-connected AI agents with security concerns that have limited broader adoption of autonomous browsing features.