dsh-omni-vision

Text-only models have no eyes, yet an Agent workflow often needs to read text from a screenshot, confirm the structure of a flowchart, or describe the color layout of an image. dsh-omni-vision gives the Agent the ability to see entirely locally, without depending on any vision model.

The plugin ships four core tools: eyes_render draws canvas elements (text, lines, rectangles, circles) in the Web GUI and renders Mermaid diagrams, saving PNG in duplicate copies; eyes_ocr calls the Windows built-in offline OCR engine to read text back into the model; eyes_analyze decodes a PNG into a pixel-color grid and dominant-color histogram so the Agent can understand shapes and palettes as structured data; eyes_paste lets the user paste or drag a clipboard image straight into the page, which the browser captures and stores as a local PNG for reading. The entire loop runs over local HTTP and disk with no unnecessary network calls.

It suits developers and teams on Windows who use DeepSeek or other text-only models and want the Agent to handle images and text locally. No extra GPU, no vision-model API, install-and-go, works offline.