dsh-plugin-vision-toolkit

In DeepSeek Harness, an agent is, at its core, a text-only reasoning engine. When a user drops a screenshot, attaches a product UI mockup, or references a chart, a pure text model has no way to engage with the visual content at all. dsh-plugin-vision-toolkit closes that gap: it plugs a pair of eyes into the agent, so image understanding becomes a natural part of the reasoning loop instead of a dead end.

The plugin ships four lightweight CLI tools, each targeting a different granularity of visual task. glance describes, answers questions about, or OCRs an entire image; ground pins down a specific element and returns its bounding box; detect scans the image for every instance of a given element type; and crop slices out a rectangular region by coordinates. Under the hood it works with DeepSeek VL (deepseek-vl2 / vl2.5), OpenAI GPT-4V / GPT-4o, or any OpenAI-compatible multimodal endpoint. The agent never touches raw pixels; it receives a natural-language description and reasons over it just like any other text.

If you are building workflows on DeepSeek Harness and your agent frequently needs to interpret screenshots, UI design comps, scanned charts, or photographs, this plugin is a drop-in fit. It does not require swapping models or restructuring your inference pipeline. Configure a single vision API endpoint, and the agent will autonomously decide when to look, which image to examine, and which tool to call.