chrome-devtools

The chrome-devtools skill provides AI agents with direct, programmatic access to a live Google Chrome instance using the Model Context Protocol (MCP). By leveraging the Chrome DevTools protocol and Puppeteer, this skill enables sophisticated browser automation, deep technical debugging, and site performance evaluation. It is designed for software engineers, QA automation specialists, and researchers who require an AI assistant to interact with complex web applications, troubleshoot network requests, or perform repetitive browser-based tasks. The agent manages a persistent browser profile, allowing for stateful sessions across multiple interactions, including support for browser extensions.

• Automated browser navigation, interaction, and content extraction using element UIDs from page snapshots.

• Advanced network request monitoring, console log analysis with source-mapped stack traces, and visual state capturing via screenshots.

• Performance tracing and site audits using built-in Chrome DevTools infrastructure.

• Extension management, including installation, trigger-based action execution, and background service worker inspection.

• Multi-page context handling with flexible selection and navigation workflows.

• Prerequisites: Requires Node.js and a current version of Google Chrome or Chrome for Testing installed locally.

• Security: The agent has the ability to inspect and modify browser data; ensure the browser instance is not used for sensitive personal accounts or credentialed sessions.

• Workflow optimization: For faster interactions, prioritize text-based take_snapshot over visual screenshots unless visual confirmation is explicitly required by the user.

• Lifecycle management: The browser initiates on the first tool call; configuration via CLI flags like --categoryExtensions or --slim (for basic tasks) allows for customized performance footprints.

• Troubleshooting: If the MCP server fails, standard Chrome DevTools UI and documentation remain the primary fallback for manual intervention. Note that parallel tool execution is supported but requires strict adherence to sequential dependency patterns: navigate, then wait, then snapshot, then interact.