Introduction

The core capability of DeepSeek Harness (DSH) lies in pluggable extension. When developers need to use z.ai’s GLM Coding Plan-related capabilities within DSH workflows, they usually need to manually integrate MCP servers. The dsh-plugin-zai-mcp plugin uses DSH’s built-in @deepseek-ai/dsh-mcp-client bridge layer to expose z.ai’s four MCP servers (vision, web reading, web search, and zread) as native DSH tools.

Plugin Scope

The plugin is maintained by the user nekomona and falls under the “Model Inference” category. It requires no build step and can be integrated into DSH’s plugin ecosystem through declarative configuration.

Core Capabilities

The plugin converts z.ai’s MCP servers into DSH tools, with 13 tools in total. Tool calls consume GLM Coding Plan quota.

  1. zai-vision
    - Transport: stdio
    - Tools:

    • analyze_image
    • analyze_video
    • ui_to_artifact
    • extract_text_from_screenshot
    • diagnose_error_screenshot
    • understand_technical_diagram
    • analyze_data_visualization
    • ui_diff_check

  2. zai-web-reader
    - Transport: streamable-http
    - Tools:

    • webReader

  3. zai-web-search
    - Transport: streamable-http
    - Tools:

    • web_search_prime

  4. zai-zread
    - Transport: streamable-http
    - Tools:

    • search_doc
    • get_repo_structure
    • read_file

Installation

The plugin supports installation from a Git source, and DSH automatically handles the dsh.profile.bundles configuration.

dsh plugin --profile web add "github:nekomona/dsh-plugin-zai-mcp"

Configuration and Key Management

The plugin requires zero configuration by default. If DSH already has a GLM Coding Plan key configured, the plugin automatically reuses that credential.

The API key resolution order is as follows:
1. config.apiKey
2. The environment variable Z_AI_API_KEY or ZAI_CODING_CN_API_KEY
3. DSH’s managed credential store

The key is used for reading only and is never written to disk or logs. To customize it, configure apiKeyEnvNames or set config.apiKey directly.

Notes

  • Cloud Dependency: All tool calls are made through z.ai / Zhipu services (open.bigmodel.cn). URLs, images, videos, and other data sent to the tools pass through these cloud services.
  • Quota Limits: Tool calls consume GLM Coding Plan quota.
  • zread Limits: zread only supports public repositories indexed by zread.ai and cannot read private or non-indexed code repositories.
  • Vision Limits: When zai-vision processes local media files, the size limit is 8MB.
  • License: The license is currently not explicitly listed (uncertain).

Summary

This plugin provides DSH users with a way to use z.ai services without manually configuring MCP clients. For more technical details and source code, refer to its GitHub repository.

  • Plugin Directory: https://www.skillhub.cn/plugins/nekomona/dsh-plugin-zai-mcp
  • GitHub: https://github.com/nekomona/dsh-plugin-zai-mcp