The core design philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” When developing with the Web GUI, connecting MCP servers, maintaining the memory store, and managing skills and slash commands usually require manually editing configuration files, without a unified entry point. dshx, as an extension suite, provides an MCP client hub (supporting stdio / Streamable HTTP / SSE) and a Web management console, while integrating skills, slash commands, project and global memory, and prompt management features.

Core Features

dshx organizes its functionality into multiple modules. Each module exposes capabilities through Model Tools while also offering visual management through the Web console.

  • MCP Hub: Supports connecting to MCP servers using the stdio, Streamable HTTP, and SSE protocols. It provides a complete state machine (connecting/connected/reconnecting/error), automatic reconnection with exponential backoff, and real-time logs (ring buffer) plus tool lists. Discovered tools are automatically bridged to the model and named mcp__<server>__<tool>. Model tools include mcp_status and mcp_restart.
  • Memory Management: Supports global memory (stored in $DSH_HOME/dshext/memory/*.md) and project memory (stored in <workspace>/.dsh/memory/*.md). Memory is injected into the system prompt before each model request. Model tools include memory_list, memory_read, memory_write, and memory_delete. Project-scoped writes follow the session sandbox policy (approval and permission elevation are required).
  • Skills and Command Management: Skills and slash commands are both managed as files. Skill files are located in $DSH_HOME/skills or <workspace>/.dsh/skills; command files are YAML and located in $DSH_HOME/dshext/commands or <workspace>/.dsh/commands. Skill changes are reflected in the session in real time. Commands are invoked using /cmd <name> and support argument replacement.
  • Prompt Management: Provides prompt library management, supports configuring prompt order and enabled status, and includes a system prompt builder.
  • Web Management Console: The plugin injects a “⚙ Extension Management” floating button into the DSH Web GUI. Clicking it opens the management interface. The console is bound to the local address http://127.0.0.1:3080/dshext/ and provides five tabs for MCP, skills, commands, memory, and prompts, along with a JSON API.

Installation and Enablement

Installing dshx requires removing old local manual configuration lines before installation to avoid duplicate registration.

  1. Run the following command to add dshx to the Web server configuration:
    dsh plugin --profile web add github:YZz-S/dshx
  1. Restart the dsh web service.

After installation, an extension management entry appears in the Web interface, and DSH automatically loads the modules described above.

Typical Usage

Configure an MCP Server

Add a server through the “MCP” tab in the Web management console. For example, add a GitHub MCP server (stdio protocol):

{
  "servers": [
    {
      "id": "github",
      "name": "GitHub",
      "transport": "stdio",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_TOKEN": "your_token_here" }
    }
  ]
}

After adding it, the model can directly call tools provided by the server, such as mcp__github__search_repos.

Define a Slash Command

Create a YAML file in the “Commands” tab of the Web management console or in an editor:

# $DSH_HOME/dshext/commands/review.yaml
name: review
description: Review the current changes
hint: '[path]'
prompt: |
  Review {{args}} and output an issue list ordered by severity.

After defining it, users can invoke the command with /review. The model tool command_define can also be used to define commands dynamically.

Applicable Scenarios and Notes

dshx is suitable for developers who need to manage multiple MCP servers, as well as long-term agent memory and capabilities, from a unified Web interface.

  • Tech Stack: Pure ESM JavaScript, no build step required, dependencies distributed with DSH installation.
  • Security: Credentials are passed only via environment variables or configuration values, with no hardcoding; the console is bound only to local 127.0.0.1.
  • Permissions: Project-scoped writes (such as writing project memory) follow the session sandbox policy and require approval.
  • Ecosystem Background: This plugin works through the DeepSeek Harness plugin mechanism. The community catalog (https://www.skillhub.cn/plugins/YZz-S/dsh-dshx) provides an index, and the plugin itself is maintained by the community.

Summary

dshx centralizes MCP server connection, memory storage, skill definition, and command dispatch through its Web console, reducing the tediousness of manual configuration. It is a practical extension for building complex agents.