Can DeepSeek Harness see images?

Not with text-only models, but vision plugins bridge the gap. modlens and dsh-vision-toolkit send images to vision models and return structured results; agent-vision-toolkit adds image Q&A, OCR and GUI automation for coding agents.

Which vision plugin for DeepSeek Harness should I use?

modlens for turning images into structured JSON evidence (OCR, layout, semantics), dsh-vision-toolkit for intent-aware Q&A and long-screenshot OCR, and agent-vision-toolkit if you want full GUI automation on top.

How do I install a vision plugin?

One command, e.g. 'dsh plugin add github:liustack/modlens'. Vision plugins typically need a compatible vision model (OpenAI-compatible VLM endpoint) configured.

How to Make DeepSeek Harness See Images: Vision Plugins Explained

Text-only models are great until you need them to look at something. The DSH community asks “how can I make DeepSeek Harness see images?” a lot — the answer is a vision plugin. This guide maps the options.

The short answer

DeepSeek Harness with a text-only model can’t see images directly. Vision plugins fix this in two ways:

Vision bridges — send the image to a vision model and get back structured results
Vision toolkits — image Q&A, OCR, screenshots, UI restoration, GUI automation

The three plugins that matter:

modlens — the vision bridge: paste an image, get structured JSON evidence
dsh-vision-toolkit — intent-aware image Q&A, long-screenshot OCR, UI restoration
agent-vision-toolkit — full vision + GUI automation for coding agents

How each one works

modlens — image to structured evidence

The most direct “make the model see” tool: paste an image into the conversation and the agent receives structured JSON — OCR, layout, semantics — that it can reason about. Zero Python, one-command install.

dsh plugin add github:liustack/modlens

dsh-vision-toolkit — intent-aware vision

Built specifically for text-only models: you can ask questions about the image, do long-screenshot OCR (great for docs and web pages), and restore UI from screenshots. It detects the intent behind each vision request and routes it appropriately.

agent-vision-toolkit — vision + GUI automation

The heavyweight: multi-image understanding, image Q&A, long-screenshot OCR, frontend UI restoration and GUI automation, plus a skill wrapper. If you want the agent to operate interfaces it sees, this is the one.

Use cases by scenario

“Look at this error screenshot and tell me what’s wrong” → modlens
“OCR this long web page / PDF screenshot” → dsh-vision-toolkit
“Rebuild this UI from the screenshot” → dsh-vision-toolkit or agent-vision-toolkit
“Click the button you can see on screen” → agent-vision-toolkit (GUI automation)

Configuration note

Vision plugins need a vision-capable model. Most accept any OpenAI-compatible VLM endpoint (for example a free-tier Chinese provider works). The plugin pages in this directory link the exact setup steps.

Browse the Tools & Capabilities category or search for vision plugins for more options.

Bottom line

If you keep hitting “I can’t see that image,” install modlens first — it’s the fastest path from text-only to “the agent can read screenshots.” Add dsh-vision-toolkit when you need OCR-at-scale or UI restoration.