jkudish/jev-browser

What it solves

Jev Browser is designed to automate browser-based tasks using an AI agent that can navigate websites, interact with elements, and extract information. It aims to provide a fast and cost-effective way to drive a real headless browser (via Playwright) to achieve specific goals, such as filling forms, pulling pricing data, or navigating through Wikipedia articles.

How it works

The project uses TypeSafe's Jev model to make decisions at each step of the navigation. The model analyzes the page's clickable, typeable, and selectable elements and selects one action per step. It also scores the likelihood that the goal has been met or if the run is stuck. The system is managed by a loop that handles budgets, recovery, and stop gates. It can be delivered as an MCP server, a CLI tool, or a library for integration into other applications.

Who it’s for

• AI Agent Developers: Those looking to integrate browser automation capabilities into their agents via MCP (Model Context Protocol) or a TypeScript library.
• Automation Engineers: Users who need to perform web scraping or form filling without writing complex, custom scripts for every site.
• Developers: Anyone needing a programmatic way to interact with the web using an AI-driven approach.

Highlights

• Multiple Integration Modes: Available as an MCP server, CLI, or library.
• Secure Credential Handling: Includes a "Password Fill" mechanism and seed cookies to handle logins without exposing secrets to the AI model.
• High Efficiency: Capable of performing tasks like navigating Wikipedia in seconds for a fraction of a cent.
• Detailed Traceability: Provides a step-by-step trace with confidence scores, final page content (including markdown), and screenshots.
• Handoff Files: Uses one-shot handoff files for secure delivery of passwords and cookies to the MCP server.