Introduction

In agent development with DeepSeek Harness (DSH), processing web content usually relies on DOM parsing, but this does not cover visual elements on the page. The dsh-browser-vision plugin provides a solution. It acts as a browser tool, allowing agents to “see” the page rather than only reading its structure.

Plugin Role

This is a browser tool plugin maintained by tristan-mcinnis. It executes tasks by driving a real Chrome browser (via Chrome DevTools Protocol, CDP) and uses a vision model to parse page content.

Core Features

The plugin has the following core capabilities:
* Page Control: Drives a real Chrome browser via CDP for operations.
* Visual Recognition: Can read text rendered on Canvas, text inside images, and rendered charts.
* Structured Extraction: Supports schema-based JSON field extraction for convenient downstream processing.
* Output Format: Supports Markdown output.
* Cost Tracking: Provides token and cost statistics for each task run.

Installation and Enablement

The installation steps are described below.

  1. Add the plugin:
    Run the following command in the terminal to add the plugin to the DeepSeek Harness web profile:
    dsh plugin --profile web add "github:tristan-mcinnis/dsh-browser-vision"
  1. Install the engine:
    Install the Python engine (using uv):
    uv tool install "git+https://github.com/tristan-mcinnis/dsh-browser-vision"

(Note: `pip` or `pipx` can also be used.)

  1. Configure the environment:
    Ensure DEEPSEEK_API_KEY is set in the shell environment. The plugin does not store or forward the key; it only inherits the runtime environment.

  2. Restart the profile:
    Restart DSH’s web profile (this profile usually has hot reload disabled).

  3. Verify status:
    After adding it, you can call the browser_vision_status tool to confirm whether the engine is available.

Typical Usage

The plugin provides three main tools:

  • browser_vision_task: Runs web tasks using natural language. It navigates, clicks, types, and reads content that exists only as pixels.
  • browser_vision_extract: Extracts page fields according to the provided schema and outputs JSON.
  • browser_vision_status: Reports engine installation status for troubleshooting installation issues.

System Requirements and Notes

  • Python Version: Python 3.11 or 3.12 is required.
  • Browser Dependency: The system must have a Chrome or Chromium binary installed. Note: this plugin does not use Playwright, so there is no need to run playwright install.
  • Visual Detail: When using the screenshot feature, avoid using the detail: low parameter because it crops text and makes it unreadable.

Conclusion

dsh-browser-vision provides basic visual processing capabilities for DeepSeek Harness, enabling agents to handle Canvas, charts, and other content that traditional DOM parsing cannot access. For web tasks that require processing complex visual information, it is a practical supplementary tool.

  • Plugin Catalog: https://www.skillhub.cn/plugins/tristan-mcinnis/dsh-browser-vision
  • Project Homepage: https://github.com/tristan-mcinnis/dsh-browser-vision