AI Agent Hub
Back to plugins
🤖

dsh-vision-bridge

Model Inference Updated 2026.08.21

Run the following command in DeepSeek Harness:

dsh plugin install zzdream67/dsh-vision-bridge

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

Run dsh plugin install zzdream67/dsh-vision-bridge in a terminal with Node.js to install the plugin from https://github.com/zzdream67/dsh-vision-bridge into DeepSeek Harness.

About this plugin

In DeepSeek Harness, pasting a screenshot into a text-only model is refused outright. dsh-vision-bridge removes that wall: you stay on the same text model, and before the request reaches the provider, every image block is swapped for a transcription produced by a vision model. No model switching, no extra route registered, nothing extra in the model picker.

The plugin intercepts the llm/stream waterfall and rewrites image blocks at any nesting depth, including those left behind inside tool-result payloads. Built-in features include transcription caching, per-image timeout, a failure notice that tells the model not to pretend it saw the image, a prompt-injection fence around all recovered text, and automatic input: [text, image] declaration management in settings.yaml that is written on enable and precisely withdrawn on disable or unload. Every field is editable live from a browser-side settings panel—no restart required.

Built for developers working with text-only models (DeepSeek, local Ollama / LM Studio, etc.) who need the model to read screenshots, UI layouts, code logs, or tables. The plugin ships no vision service of its own; you supply the vision model—a local instance keeps data on your machine at zero cost, or a remote API billed to your own account.

Use Cases

  • Let a text-only model read errors and logs from a screenshot
  • Analyze UI layout and button positions without switching models
  • Convert tables and chart trends in an image into text for the text model to reason over

Best For

  • Developers using text-only models in DeepSeek Harness
  • Individuals pairing a local Ollama or LM Studio vision model
  • Users who need image reading without switching away from their text model