alchaincyf/huashu-chrome

What it solves

It addresses the limitation where AI agents typically operate in isolated, "clean" browser sessions without access to your existing login states, cookies, or personal browser history. It allows any MCP-compatible AI agent to control your actual, daily-use Chrome browser, enabling tasks like filling forms, interacting with authenticated dashboards, or scraping data from sites where you are already logged in.

How it works

It functions as a bridge between an AI agent and your browser using a five-part architecture: an npm package, a CLI, an MCP server, a local bridge process, and a Chrome extension. When you issue a command, the agent communicates via the MCP server to the local bridge, which then relays instructions to the Chrome extension. The extension acts as the "hand" that interacts with the browser, providing the agent with structured snapshots of the page (using ref-numbered elements) and executing actions. It includes built-in "site knowledge" for over 20 websites to help the agent navigate common pitfalls, and it allows the agent to learn and save site-specific patterns locally.

Who it’s for

This project is for power users and developers who use AI coding agents (like Claude Code, Cursor, Cline, or Windsurf) and want to automate browser-based workflows that require authentication, such as CRM data entry, managing cloud consoles, or interacting with internal web tools.

Highlights

• Real Login Persistence: Operates directly within your active Chrome browser, meaning no need to re-authenticate or handle complex login flows.
• Verification-First Actions: Every write operation requires the agent to report evidence of page changes; if a page does not react as expected, the agent stops to prevent cascading errors.
• Site-Specific Intelligence: Comes with pre-configured "experience" notes for over 20 popular sites and allows the agent to save new findings locally to improve future performance.
• Safety-Focused Design: Implements security measures outside the LLM, such as a mandatory confirmation UI for payments that the agent cannot bypass, and automatic sensitive data redaction.
• Broad Compatibility: Automatically detects and configures MCP settings for over 20 popular AI agents and CLI tools.