Foreword

In the development of DeepSeek Harness (DSH), enabling agents to operate a browser is a common requirement. Existing general-purpose solutions often require launching a separate browser profile or performing low-level DevTools protocol debugging. This usually causes loss of user state, makes the interaction process invisible, and makes fine-grained ownership management difficult.

dsh-native-browser aims to solve the above problems. It does not attempt to become a general-purpose Chrome MCP server. Instead, it focuses on DSH’s product-layer control: operating the Chrome browser the user is currently using, preserving logged-in sessions and extensions, and providing a visual handoff mechanism.

What It Is

This is a Manifest V3-based Chrome extension that communicates with a local runtime via Native Messaging. It provides DSH with a native browser runtime, allowing agents to operate tabs in the user’s existing profile without launching a second browser instance.

Core Features

The plugin provides the following core capabilities:

  1. Semantic observation and low latency
    It uses an accessibility-tree-first observation strategy and provides compact incremental updates. Screenshots are used only as a visual fallback, not as the primary source of structural observation.

  2. Reliable interaction and verification
    It provides control based on CDP (Chrome DevTools Protocol), but verifies the hit target before input. Persistent operation logs ensure that the semantic result of an operation is “verified” or “unknown,” rather than blindly assuming input success.

  3. Explicit ownership management
    It has clear ownership of existing tabs and tabs created by the agent. Temporary tabs created by the agent are automatically cleaned up after the task ends.

  4. Visual interaction and handoff
    It provides display-only virtual pointers and pulse animations for clicks/scrolling, enabling human users to perceive the agent’s operations. When the DSH UI switches conversations or the agent releases control, permissions are automatically revoked. Humans can interrupt the agent’s operations at any time.

Installation and Enablement

Because an official one-command installation is not yet provided, the environment currently needs to be configured manually.

  1. Prepare the DSH environment
    Ensure that DeepSeek Harness version 0.1.5-rc.1 is installed.

  2. Load the Chrome extension
    After obtaining the code from the project repository, open the Chrome “Extensions” management page in the Chrome browser, choose “Load unpacked extension,” and select the plugin directory.

  3. Configure Native Messaging
    Configure the Native Messaging Host according to the project documentation to establish the communication channel between the local runtime and the extension.

Notes

  • Current status: It is in Public Alpha and is not recommended for direct use in production environments.
  • Browser support: Chrome is currently prioritized. Edge compatibility is in early validation (architecture gate tests only).
  • Page compatibility: Extensive page compatibility and production hardening work has not yet been completed.
  • License: It follows the MIT license.

Catalog and Source Code

Brief Closing

dsh-native-browser provides a transition from a “debugging tool” to a “product control layer,” making it suitable for developers who want to achieve controllable and observable interactions in an existing browser environment.