AI Agent Hub
Back to plugins
🤖

dsh-vision

Model Inference Updated 2026.08.15

Run the following command in DeepSeek Harness:

dsh plugin install joyiok/dsh-vision

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

Run dsh plugin install joyiok/dsh-vision in DeepSeek Harness to install this vision plugin; the source repository is available at https://github.com/joyiok/dsh-vision. Restart dsh web after installation to activate it.

About this plugin

Your DeepSeek Harness agent is chugging along when it hits a screenshot, a design mockup, or a crash log rendered as an image, and your text-only main model simply cannot see it. That is the gap dsh-vision fills: it bolts a pair of eyes onto any non-multimodal model so the agent can finally interpret visual content.

Once installed, the agent gains an image_analyze tool. You pass in local file paths, http(s) URLs, or data URLs (up to four images), and the plugin base64-encodes them in memory, forwards them to an OpenAI-compatible vision endpoint, and hands the textual analysis back to the main model for continued reasoning. The default backend is a Gemini Flash model via OpenRouter, supporting png, jpeg, gif, and webp. You can override the model, cap output tokens, or frame a specific question — for instance, "check whether this UI has alignment issues." Built-in retry logic handles 429 and 5xx responses, and AbortSignal-based timeouts let the upper layer cancel requests cleanly.

If you are running a text-only LLM such as DeepSeek or Llama as the Harness brain and want the agent to develop a working ability to read screenshots, scan dashboards, and review interfaces, dsh-vision is a lightweight, low-friction way to close that capability gap.

Use Cases

  • Calling a vision model to extract text from a screenshot when the main model cannot read it
  • Interpreting UI screenshots, dashboard images, or error logs in automated workflows
  • Understanding a pasted image URL or local file so the agent can proceed with the next action

Best For

  • Developers using text-only LLMs like DeepSeek or Llama as the Harness brain
  • Automation engineers who want basic image recognition inside their agent pipelines
  • Users who need a lightweight vision add-on without migrating to a multimodal main model