Preface

The DSH plugin system emphasizes “everything is a plugin.” The community directory is an independent site and has no official affiliation with DeepSeek / Hypersphere. For text models using DeepSeek Harness, a specific limitation is that inputModalities does not contain “image,” so messages containing images will be rejected. dsh-image-mmx addresses this issue: after pasting or sending an image in the session, it automatically calls mmx (MiniMax VLM) to recognize it and injects the recognition result into the model context.

What is this

dsh-image-mmx is a DeepSeek Harness plugin maintained by fengs2021, with the repository path fengs2021/dsh-image-mmx and an MIT license.

One-sentence positioning: Gives text models in DeepSeek Harness a pair of eyes: paste/send images in the session, automatically call mmx (MiniMax VLM) to recognize, and inject the recognition result into the model context — pure text models like DeepSeek can now understand image content.

Core Features

Below are the verified capabilities:

  • Text models (where inputModalities has no image) can send messages containing images without reporting “Current model does not support images”
  • Images are saved to the session workspace .attachments/, automatically executing mmx vision describe (MiniMax VLM)
  • Image blocks are replaced in the model’s view with [图片N]:"<path>" + mmx recognition text (preserving scene elements, layout, error/code/interface text verbatim)
  • Human transcripts render original image thumbnails as usual (clickable to enlarge)
  • Falls back to path text when mmx is not installed/unauthenticated/timeout (60s)/unsupported format (gif)
  • Models that natively support image input take the native path without any processing
  • Multi-image support: Each image is recognized independently and in parallel, numbered as [图片1] [图片2]...

Installation and Activation

First, confirm that the environment meets the following prerequisites:

  • DeepSeek Harness 0.1.0-rc.6+
  • Node.js 18+
  • peerDependencies: @deepseek-ai/cordis ^4.0.1
  • mmx-cli (MiniMax official CLI) installed and authenticated locally

If mmx is not installed, unauthenticated, recognition times out (60s), or the image format is unsupported (gif), the plugin will fall back to path text.

Installation command:

dsh plugin --profile web add 'github:fengs2021/dsh-image-mmx'

Typical Usage

  1. Paste or drag images into the DSH Web input box, or select images using the attachment button from dsh-file-bridge.

  2. After sending, the plugin automatically completes the process: saving the image, calling mmx for recognition, and the model directly reads the recognition result to reply.

  3. With multiple images, each image is recognized independently and in parallel, numbered as [图片1] [图片2]....

Applicable Scenarios and Notes

Suitable for: DSH Web sessions using pure text models like DeepSeek that need to convert screenshots, errors, code, or interface content into text context readable by the model.

Notes:

  • The plugin runs with the permissions of the current dsh process; check the source code and license before installing.
  • Models that natively support image input will take the native path without extra processing.
  • For unsupported formats like gif, mmx not installed/unauthenticated, or timeout (60s), it will fall back to path text.
  • The community directory is an independent site and has no official affiliation with DeepSeek / Hypersphere.

Conclusion

The value of dsh-image-mmx lies in: users still see image thumbnails, while the model side reads mmx’s recognition text, allowing text models to continue replying based on image content.

Directory Page: https://www.skillhub.cn/plugins/fengs2021/dsh-image-mmx

GitHub: https://github.com/fengs2021/dsh-image-mmx