dsh-vision-bridge

In DeepSeek Harness, many models are text-first and lack native image understanding capabilities, which limits users from performing image-related tasks like visual question answering or image analysis. The dsh-vision-bridge plugin bridges this gap by adding a vision layer to text models without altering the models themselves, thus expanding their functional range.

Its core abilities include integrating multiple vision providers (such as Gemini, OpenAI-compatible, and Anthropic-compatible APIs) and offering a per-model glasses toggle to enable or disable image understanding. It securely processes images by sending them to the configured vision provider and returning textual analysis, while keeping the model list clean. Additionally, it supports persistent preferences to ensure user choices remain consistent across sessions and provides clear information about the active vision provider.

It is ideal for users of DeepSeek Harness who need image analysis capabilities, such as developers building multimodal applications, researchers conducting visual experiments, or everyday users looking to extend the utility of text models. It simplifies the process of adding visual abilities to existing models without requiring complex configuration or model switching, making image understanding more accessible and efficient.