Introduction

If you need a DSH agent to complete web browsing, clicking, form filling, and screenshotting, a common approach is to add an extra layer on top of CDP endpoints or MCP services. The documentation for duyefeng/dsh-browser explicitly states: no CDP configuration needed, and no MCP service running required.

It uses Playwright under the hood to directly drive the locally installed Microsoft Edge, encapsulating browser actions into a set of browser_* tools for the agent to call.

What is this

dsh-browser is a browser plugin for DeepSeek Harness (DSH). Once installed, the AI can directly open web pages, click buttons, fill out forms, and take screenshots.

The repository is under the duyefeng account: https://github.com/duyefeng/dsh-browser. The license is MIT.

Core Capabilities

  • Built on Playwright, directly driving the Microsoft Edge already installed on the computer.
  • Page behavior mimics human browsing; cookies, JS, and login states are preserved.
  • No separate CDP configuration is needed, and no MCP service needs to be run.
  • Provides the following browser tools to the agent:
Tool Description
browser_navigate(url) Open URL
browser_snapshot() Read current page, e.g., title, body, clickable elements
browser_click(selector) Click an element
browser_type(selector, text) Type text into an input field
browser_press(key) Press a key, e.g., Enter, Tab, Esc
browser_evaluate(expression) Execute JavaScript within the page
browser_screenshot() Take a screenshot and return the image path
browser_close() Close the browser for the current session

package.json declares a dependency on playwright-core ^1.49.0 and peerDependencies: @deepseek-ai/cordis, @deepseek-ai/dsh-tools.

Requirements

Before using this plugin, confirm that the environment meets the following:

  • Node.js 22+.
  • Microsoft Edge is installed.
  • You have the dsh command line tool; if running from a source code checkout, you can use pnpm dsh.

Installation and Activation

First, install the plugin:

dsh plugin --profile web add github:duyefeng/dsh-browser

Then start:

dsh web

The plugin is read when the profile is started. After installation, you need to restart the profile for it to take effect.

Typical Usage

After installation, you can directly use natural language to instruct the AI to perform browser tasks, for example:

  • “Open example.com, read the body, and tell me the title.”
  • “Open Baidu to search for ‘weather’ and give me the title of the first result.”
  • “Open this login page, fill in the account and password, and take a screenshot for me to see.”

The AI will call the corresponding browser_* tools based on the task.

Override Default Configuration

If the default behavior does not meet requirements, you can override the browser configuration in the profile’s cordis.patch.yml:

- id: browser
  config:
    channel: msedge      # Use the installed Microsoft Edge
    headless: false      # false: with UI; true: headless
    timeoutMs: 30000     # Timeout for a single operation, in milliseconds

Use Cases and Notes

Suitable for scenarios where web browsing, clicking, form filling, and screenshotting need to be completed within DSH, and Microsoft Edge is already installed on the local machine.

Because it directly operates the browser and reads page content, it is a sensitive capability. Before installation, it is recommended to check the source code and the MIT license; when running, it will execute browser operations with the permissions of the current dsh process, so please ensure the connected pages and input content are trustworthy.

Links

  • GitHub: https://github.com/duyefeng/dsh-browser
  • Directory Page: Search for dsh-browser in the DSH plugin directory where it is located; the community directory is independent of DeepSeek / Hypersphere and is not an official app store.