Browse

Browse

Read and interact with Chrome browser pages using DOM-based tools.

Workflow

1. Get tab context

Always start by calling mcp__claude-in-chrome__tabs_context_mcp to see existing browser tabs.

2. Open a page

• New tab: mcp__claude-in-chrome__tabs_create_mcp with the target URL
• Existing tab: mcp__claude-in-chrome__navigate to change the URL of a known tab

If the user provided a URL argument, open it now.

3. Read page content

• Primary: mcp__claude-in-chrome__read_page -- returns the accessibility tree of DOM elements (structured, interactive elements visible)
• Text extraction: mcp__claude-in-chrome__get_page_text -- returns extracted text content, prioritizes article/main content

Use read_page first. Fall back to get_page_text if you need cleaner prose (e.g., reading an article or documentation page).

4. Find and interact with elements

• Find elements: mcp__claude-in-chrome__find -- use natural language queries to locate elements (e.g., "login button", "email input field")
• Fill forms: mcp__claude-in-chrome__form_input -- use element reference IDs returned by find to fill inputs, click buttons, select options
• Run JavaScript: mcp__claude-in-chrome__javascript_tool -- execute custom DOM queries or page manipulation when the other tools are insufficient

5. Debug

• Console logs: mcp__claude-in-chrome__read_console_messages -- read browser console output (use pattern parameter to filter)
• Network requests: mcp__claude-in-chrome__read_network_requests -- inspect network traffic for API calls, errors, etc.

6. Record interactions (optional)

Use mcp__claude-in-chrome__gif_creator to record multi-step browser interactions as a GIF for the user to review.

Critical Rules

1. NEVER use mcp__claude-in-chrome__computer for reading pages. Always use read_page and get_page_text for DOM-based content extraction. The computer tool uses screenshots which are slower, less structured, and more token-expensive.

2. Never trigger JavaScript alerts, confirms, or prompts. These browser dialogs block all further browser events and prevent the extension from receiving commands. Use console.log + read_console_messages for debugging instead.

3. Bail out after 2-3 failures. If browser tools fail or return errors after 2-3 attempts, stop and ask the user for guidance. Do not loop or retry indefinitely. Explain what you tried and what went wrong.

4. Never reuse tab IDs from a previous session. Always call tabs_context_mcp to get fresh tab IDs if there's any doubt.