Introduction

In the DeepSeek Harness (DSH) ecosystem, agents handling complex web interactions—such as preserving login state, solving captchas, and dealing with dynamic rendering—often struggle with command-line tools alone. General-purpose browsers are designed for humans and are difficult for agents to control precisely. dsh-browser-cdp (formerly dsh-ego-browser) connects to a real Chrome browser through Chrome DevTools Protocol (CDP), provides agents with 38 structured tools, and includes a real-time observation window, solving the problem of agents in browsers being “unable to see or intervene.”

Core Capabilities

This plugin provides the following core features:

  1. 38 structured tools: A set of bcdp_* tools covering concrete actions such as page operations, screenshots, and downloads, with single responsibilities and high determinism.
  2. Real-time observation window: Streams the page view in real time via SSE, supports direct mouse control of the real browser and download capture from the observation window. If dsh-better-sidebar is installed, the observation window is displayed as a sidebar tab.
  3. Login-state import: Supports copying login cookies from Chrome/Edge/Brave into the agent’s browser by domain, compatible with Chrome 127+ App-Bound Encryption.
  4. Self-observation: When an agent operates DSH’s own interface (such as managing sessions), its actions are also visible through the observation window, allowing developers to take over at any time.
  5. Platform adaptation: Automatically detects Linux, macOS, and Windows environments; includes a built-in --no-sandbox wrapper to support root/Docker/no-display environments.
  6. Operations tools: Provides auxiliary tools such as bcdp_doctor (environment health check), bcdp_captcha (captcha/human verification detection), and bcdp_http (browser-context requests).

Installation and Prerequisites

Before use, ensure your environment meets the following requirements:

  • Node.js: version >= 22
  • DSH: version >= 0.2.0-rc.1
  • Optional dependency: dsh-better-sidebar >= 0.12.2 (for sidebar support)

Install the plugin:

dsh plugin --profile web add github:drscrewdriver/dsh-browser-cdp

The plugin package includes a prebuilt runtime, so no additional build scripts are required.

Typical Usage

This plugin is suitable for the following scenarios:

  1. Literature and data scraping: Have the agent log in to CNKI or Google Scholar, collect data by paging through results, and let developers observe scraping progress in real time.
  2. Forms and login: If the agent encounters a captcha while filling out a form, developers can manually take over in the observation window, complete verification, and then hand control back to the agent.
  3. QA/smoke testing: Simulate user operations on the product, using the observation window as a screen recording and interactive terminal.
  4. Self-observation: Observe how the agent operates DSH’s own interface, making it easier to debug interaction logic.

Notes

  • Headless mode: Even if the observation window is not started, bcdp_* tools are still available in headless sessions, but the page view is not displayed.
  • Compatibility: Cookie import is compatible with Chrome 127+ App-Bound Encryption and does not require offline decryption.
  • Permissions: The plugin runs with the permissions of the current DSH process. It is recommended to review the source code and license before installation.

Summary

dsh-browser-cdp turns a “black-box” browser into a “white-box” observation and control terminal, making the agent’s actions in the browser fully visible and controllable. For developers who need to handle complex web interactions, it provides the necessary control levers.