Lum1104/dsh-browser

What it solves

This project provides a bridge between the DeepSeek Harness (dsh) agent framework and a live Google Chrome browser. It allows an AI agent to interact with the web as a real user would, using the user's existing browser session, login states, and cookies, rather than operating in a headless browser environment.

How it works

The system consists of a Chrome MV3 extension and a bridge plugin for dsh. Instead of using vision-based processing, it converts web pages into structured text with a numbered inventory of interactive elements. The AI model interacts with these elements by referring to their numbers to perform actions like clicking, typing, and scrolling.

To ensure privacy and security, sensitive data like passwords and credit card numbers are masked before they leave the page. The extension also requires user approval for actions like navigation and typing to prevent unauthorized control.

Who it’s for

Developers and users of the DeepSeek Harness who want to automate browser-based tasks using an AI agent that can operate within their active, authenticated browser sessions.

Highlights

• Session Preservation: Works in the user's actual browser tab, retaining all active logins and cookies.
• Text-First Interface: Uses structured text and numbered controls instead of screenshots, reducing model overhead and improving speed.
• Privacy-Centric: Masks sensitive form fields and uses a guarded bridge with bearer-token authentication.
• Comprehensive Toolset: Includes tools for reading page snapshots, clicking elements, filling forms, pressing keys, and navigating history.
• High Performance: Benchmarked to be faster and require fewer tool calls than a Playwright-based baseline.