Introduction

When building agents in DeepSeek Harness (DSH), a common requirement is that models run on pure text routing, but tasks need to handle images, videos, documents, code, and data. DSH allows extending session capabilities via plugins; dsh-qwen-mm integrates the capabilities of Qwen-MM-Plugins into DSH and provides image attachment bridging, allowing pure text models to read multimodal content via MCP tools.

What is this

dsh-qwen-mm is a DSH plugin repository maintained by RRRosmontis, licensed under MIT. It targets developers who wish to use multimodal reading, cloud vision, search, video tools, and 3D/CAD tools within DeepSeek Harness.

The plugin provides three categories of content:

  • Built-in multimodal skills;
  • MCP server rows;
  • Image attachment bridging for pure text model routing.

Core Features

Multimodal Reading

Supports reading images, videos, documents, code, and data. The core capability does not require a key.

Cloud Vision, OCR, ASR, and Object Localization

Provides capabilities such as cloud vision, OCR, ASR, and object localization. The related api capability requires a DashScope key.

Provides web search, web page extraction, and image search. The search capability requires Serper, Exa, or Tavily.

Video and 3D/CAD Capabilities

Supports long video Q&A, video editing, Blender, FreeCAD, and math video related tasks.

video-memory, video-edit, edu-agent require DashScope keys; video tools require ffmpeg; Blender, FreeCAD, LibreOffice, and Chromium are only used on demand by the capabilities that call them.

Drag-and-Drop Image Upload

You can drag images into the input box, which will be read by the pure text model via tools. This feature relies on core changes to DeepSeek Harness and has not yet entered the official release version. You need to use the RRRosmontis/deepseek-harness fork containing these changes, or apply deepseek-harness-core.patch.

Installation and Usage

Install Plugin

If you only want to install the plugin itself, you can use the following command:

pnpm dsh --profile web plugin add github:RRRosmontis/dsh-qwen-mm

After installation, you need to restart the DSH web profile and open a new session.

Prerequisite for Drag-and-Drop

The drag-and-drop feature requires additional processing for core-side changes. Optional ways are:

  1. Use the RRRosmontis/deepseek-harness fork containing these changes;
  2. Apply deepseek-harness-core.patch and then use DSH.

This patch was generated based on the DeepSeek Harness master tip (commit 47f9438). For older or newer versions, manual merging may be required.

Keys and Dependencies

DSH filters out credential-like variables in the MCP subprocess environment, so keys should be written to a shared configuration file rather than environment variables.

Distinguished by capability:

  • core: no key required;
  • api, video-memory, video-edit, edu-agent: requires DashScope key;
  • search: requires Serper, Exa, or Tavily;
  • Video tools: require ffmpeg;
  • Blender, FreeCAD, LibreOffice, Chromium: only used on demand by the capabilities that call them.

Typical Usage

Propose a task after referencing files in a session:

@report.pdf    Summarize page 3 and extract tables.
@meeting.mp4   Transcribe this meeting and annotate speakers and timestamps.
@place.jpg     Identify the location where this photo was taken and verify it online.

You can also directly drag images into the input box to ask questions. Note: The drag-and-drop feature requires using the fork or core patch mentioned above.

Suitable Scenarios and Notes

Suitable for developers who want to add multimodal input, search, video processing, or 3D/CAD toolchains to pure text model routing in DSH. Before use, you need to confirm:

  • Whether drag-and-drop is needed; if yes, handle the fork or core patch first;
  • Whether the current profile is configured with the corresponding keys;
  • Whether the local machine already has ffmpeg, Blender, FreeCAD, LibreOffice, Chromium, etc., as needed dependencies;
  • The content of assets/skills/ comes from Qwen-MM-Plugins licensed under Apache-2.0, and the upstream retains its copyright and license; the plugin itself is licensed under MIT.

DSH plugins run with the permissions of the current dsh process. Before installation, it is recommended to check the plugin source code, dependencies, core patch, and license before deciding whether to enable it.

Getting

GitHub Repository:

https://github.com/RRRosmontis/dsh-qwen-mm

The DeepSeek Harness Community Directory is a standalone site with no official affiliation to DeepSeek / Hypi; verified materials do not provide the directory page URL for this plugin, so this article does not concatenate a specific address here.