Claude Code Gets a Built-In Browser for Agentic Testing

A Browser Built Into the AI Coding Assistant

Anthropic shipped a tabbed, sandboxed browser inside the Claude Code desktop application as part of its Week 28 release, covering versions v2.1.202 through v2.1.206 during 6 to 10 July 2026. The feature is a notable first for AI coding assistants at production scale: rather than relying on a developer to manually navigate to a live web application, observe its rendered state, and describe what they saw back to a chat window, Claude Code can now open a browser pane directly inside the desktop app, navigate to any URL, and interact with the page the same way it interacts with a local development server preview.

The browser is accessible via Cmd+Shift+B on macOS or Ctrl+Shift+B on Windows, or from the Views menu in the desktop application. It is available only to paid Claude users on the desktop app — it is not part of the free tier, and it is not present in the Claude.ai web interface or mobile applications.

What Claude Code Can Do in the Browser

The in-app browser gives Claude Code access to the runtime state of a web page, not merely its source code. The agent can read page content, inspect the Document Object Model, click interface elements, fill in input fields and forms, submit those forms, and capture screenshots of the viewport at any point in a session.

These are the same actions a human QA engineer performs when manually testing a web application. Giving them to Claude Code means the agent can observe a component's rendered behaviour rather than inferring it from static source code alone — a meaningful distinction, because what a browser renders from a React or Next.js component often differs from what the source naively suggests, particularly when styles, hydration, and runtime state are involved.

The browser also supports Google OAuth pop-ups and similar identity flows, which matters for teams building SaaS applications. Authenticated features of a web app — those behind a login gate — can now be tested by Claude Code without requiring the developer to manually pass session tokens or cookies to the agent.

Sandboxing, Session Persistence, and Safety Controls

The browser runs in a sandboxed environment. Users can configure whether browsing sessions persist between Claude Code invocations — useful when returning to a partially completed multi-step workflow — or whether each session starts fresh. When Claude Code interacts with external web sites, safety classifiers review the actions the agent proposes before they are executed.

This classifier layer is an important design choice. It means the built-in browser is not a raw automation executor: it is a development and testing tool with runtime checks on what the agent can do on external domains. The safety layer limits the risk of unintended side effects when an agentic workflow visits a site the developer did not specifically authorise.

The /doctor Upgrade in the Same Release

Week 28 also upgraded the /doctor command from a read-only diagnostic report to an interactive repair tool. Running /doctor in a session now checks installation health, identifies unused skills, MCP servers, and plugins relative to their context cost, deduplicates local CLAUDE.md files against checked-in versions, and proposes trimming content that Claude could derive from the codebase itself. It presents all findings before making any changes and requests confirmation before acting. The alias /checkup also works.

How the Browser Changes the Developer Feedback Loop

Before the built-in browser, an agentic UI development cycle with Claude Code required a manual observation step: describe a bug or feature, let the agent make code changes, build and serve the application, open a separate browser, observe the rendered result, describe what was wrong, and iterate. The browser was always a step that required a human to be at the keyboard.

The built-in browser removes that mediation for a defined scope of testing tasks. Claude Code can now serve a local application, navigate to the relevant page, inspect its rendered state, identify issues, and propose a code change — all within a single agentic loop. For web scraping tasks, the change is equally direct: the agent can visit a live URL and generate extraction logic from observed DOM reality rather than from source code that may differ from what the browser actually renders.

What This Means for Development Teams in India

For Indian software product teams building web applications for domestic or international clients, the most immediate value is in visual regression verification. After a code change, Claude Code can autonomously open the affected page, screenshot the rendered result, and flag layout regressions — without requiring an engineer to be at the keyboard. For teams working across time zones with client review cycles, this compresses the turnaround on UI issues significantly.

The OAuth support is particularly practical for teams building SaaS products with Google Workspace login flows. Testing the authentication boundary of a feature previously required either a human tester with valid credentials or a complex test fixture. Claude Code can now handle that step natively, provided the appropriate paid plan is active.

The Bottom Line

Anthropic shipped a sandboxed, tabbed built-in browser for Claude Code desktop in its Week 28 release (6–10 July 2026, versions v2.1.202 to v2.1.206). Available to paid users only, the browser lets Claude Code's AI agent read page content, inspect the DOM, click elements, fill and submit forms, capture screenshots, and complete Google OAuth login flows — all within a configurable sandboxed environment where safety classifiers review actions on external sites. The in-app browser eliminates the manual observation step in agentic UI development cycles and enables direct web scraping from live-rendered pages. The same release upgraded /doctor to an interactive repair command for Claude Code installation and configuration issues.