Fisfzy/dsh-ego-browser

What it solves

General-purpose browsers are not designed for AI agents, yet many web tasks require a real browser to handle logins, captchas, and dynamic content. This project bridges the gap by connecting the ego-lite browser engine to the DeepSeek Harness (DSH) environment. It solves the "black box" problem where agents perform tasks invisibly, allowing you to see what the agent is doing in real-time and intervene manually when necessary.

How it works

It provides 33 structured ego_* tools that allow an agent to control a Chromium-based browser. When an agent initiates a task, the plugin automatically opens an observation window—either in the DSH native sidebar, a dedicated tab, or a floating bubble. This window uses real-time streaming to show exactly what the agent sees. You can click, scroll, or type directly into this window to take control of the browser session without interrupting the agent's underlying process.

Who it’s for

This tool is for developers and power users who need AI agents to perform complex web tasks—such as data scraping, form filling, or automated testing—while maintaining full visibility and the ability to manually override the agent during critical moments like solving captchas or navigating multi-step logins.

Highlights

• Real-time Observation & Control: Watch the agent work via a live stream and take manual control (click/drag/scroll) of the browser at any time.
• 33 Structured Tools: Offers granular control over browser tasks, from navigation and element interaction to executing custom JavaScript and CDP commands.
• Platform Versatility: Works across Linux, Windows, and macOS with automatic browser detection and a "headless" fallback for servers or Docker environments.
• Persistent Sessions: Supports importing cookies from your system browser and persistent profile storage, allowing agents to maintain login states across restarts.
• Self-Observation: The agent can even be observed while it interacts with the DSH interface itself.
• Robustness: Includes worker process management to handle crashes, automatic retries, and an optional "resident" mode for near-instant task execution.