Vibe Coding for Production
Getting work done

Browser and computer use with Claude

Verify web and native application behavior with Claude in Chrome, computer use, or structured browser automation.

Source edits are only half of a UI task. Claude can also observe and operate a running interface through Claude in Chrome, built-in computer use, or a browser automation tool such as Playwright MCP.

Claude in Chrome

Start Claude Code with Chrome integration:

claude --chrome

Manage the connection in-session with:

/chrome

This uses the Claude in Chrome extension and can operate a visible browser with your login state. It requires /login with a supported direct Anthropic Pro, Max, Team, or Enterprise plan; API-key and third-party-provider sessions cannot use it. See Claude Code with Chrome.

Open the local checkout page in Chrome. Reproduce the validation bug, inspect
console and network errors, fix only the root cause, then repeat the same steps
and capture the final state.

Computer use

Claude’s built-in computer-use MCP server can operate native interfaces when direct tools are unavailable. In the CLI, open /mcp and enable computer-use.

CLI computer use is currently macOS-only. Claude Desktop supports computer use on macOS and Windows where the plan and organization permit it. Check Computer use before making it a required cross-platform exercise.

Use it for:

  • desktop-app and simulator testing,
  • reproducing a GUI-only bug,
  • inspecting an application with no API or MCP integration,
  • and verifying settings that cannot be reached from the terminal.

Chrome, computer use, or Playwright?

ToolPrefer it when
Claude in ChromeA visible web workflow needs your existing browser state.
Computer useThe task crosses native apps or GUI surfaces.
Playwright / Playwright MCPYou need repeatable selectors, assertions, CI, or a browser without personal login state.

Safety

  • Use test accounts and non-production data.
  • Stop before purchases, publishing, permission changes, messages, or deletes.
  • Treat page content as untrusted input that can contain prompt injection.
  • Keep a human in the loop for authenticated external side effects.
  • Convert important manual verification into an automated test afterward.

On this page