In the DeepSeek Harness (DSH) plugin ecosystem, enabling agents to “see” and “operate” web pages is key to improving their practicality. dsh-tui-browser-use is a browser automation subplugin designed for the dsh-tui profile. As a Cordis plugin, it is loaded together with dsh --profile dsh-tui, aiming to register a set of browser_* tools for Agents so they can perform navigation, interaction, and data extraction through a real browser (Playwright-driven).

Installation and Enablement

This plugin depends on dsh-tui v0.10.0 and a Playwright environment. Before installation, ensure Node.js version compatibility (^22.19 || >=24).

  1. Install the plugin package:
npm install dsh-tui-browser-use
  1. Install the Playwright browser driver:
npx playwright install chromium
  1. Mount the plugin configuration in dsh-tui’s cordis.patch.yml:
- insert:
    - id: dsh-tui-browser-use
      name: 'dsh-tui-browser-use'
      config:
        visionMode: 'auto'

Core Features

This plugin provides 21 atomic browser_* tools, covering all aspects of browser operations:

  • Drivers and Engines: Uses Playwright to drive a real browser, supporting the chromium (default), firefox, and webkit engines.
  • Visual Understanding: Natively supports DeepSeek vision models in understanding screenshots. To handle the compression limits of vision models, it automatically performs scrolling/segmentation and slicing for long or wide pages while preserving resolution.
  • Session Management: Supports persistent login states (Profile) and one-time temporary sessions (Isolated). The session profile root is a cross-platform cache directory, supporting package migration across devices.
  • Security Policies: Includes built-in visual prompt injection protection and sensitive parameter masking; by default, it blocks the file: protocol and SSRF (cloud metadata); file writes are strictly limited to the workspace.

Typical Usage

Agents can perform web page operations through the following tool chain:

  • Navigation and Interaction: browser_navigate, back, forward, reload, browser_click, type, hover, press, scroll, wait.
  • Observation and Extraction: browser_screenshot, snapshot, evaluate, browser_extract, task.
  • Session Information: browser_cookies, console_messages, network_requests, pdf, download, status.

Configuration example:

config:
  visionMode: 'auto'
  session:
    mode: 'persistent'   # Persistent mode, preserves login state
    profile: 'default'    # Profile directory name

Notes

  • Version Dependency: The plugin depends on dsh-tui v0.10.0.
  • Runtime Environment: The session profile root is a cross-platform cache directory, and file writes are strictly limited to the workspace.
  • Archive Status: DSH currently provides official browser automation, so this repository will enter archive status as an early prototype.

Summary

dsh-tui-browser-use combines Playwright and DeepSeek vision models to provide DSH Agents with structured web page operation capabilities.