AI Agent Hub
Back to plugins
dsh-image-dscriber preview

dsh-image-dscriber

Model Inference Updated 2026.08.18

Run the following command in DeepSeek Harness:

dsh plugin install shuxu-zhao/dsh-image-dscriber

Paste the following prompt into your AI chat to install this plugin:

Run dsh plugin install shuxu-zhao/dsh-image-dsubscriber in the terminal to install the plugin, source available at https://github.com/shuxu-zhao/dsh-image-dsubscriber, then link it to your DeepSeek Harness Web Profile and declare the image modality in settings.yaml as described in the README to enable visual understanding for text-only models. Source: https://github.com/shuxu-zhao/dsh-image-dscriber

About this plugin

Text-only models in DeepSeek Harness (such as deepseek-v4-flash and deepseek-v4-pro) have no visual channel. When a user pastes a screenshot, the model either crashes on binary data or kicks off pointless local file searches, completely unable to understand the image. dsh-image-describer closes that gap.

The plugin offers dual-mode visual understanding. In attachment mode, the image is processed by a vision model (MiniMax-M3 by default) for OCR and layout recognition, then replaced with a structured prompt so the text model can answer directly from the extracted content. In tool mode, the model autonomously invokes the describe_image tool to read a workspace image file; the vision model returns a targeted analysis, and the entire call is visible in the Web UI tool card.

Built for developers who use text-only models in DeepSeek Harness but still need the model to see screenshots, architecture diagrams, error captures, and other images. Simply declare the image modality in settings and point to a vision model for visual understanding with zero extra code.

Screenshots

Use Cases

  • Inspect screenshots, error captures, or UI snapshots in a text-only model
  • Let the model autonomously analyze local architecture diagrams or flowcharts
  • Perform OCR and layout-content recognition on pasted images

Best For

  • Developers using text-only models like deepseek-v4-flash in DeepSeek Harness
  • Teams that need LLM image understanding without a dedicated multimodal model
  • Users who want to add visual understanding to existing workflows with zero code