dsh-visibridge

When working with pure-text LLMs like DeepSeek or GLM, a persistent pain point is simple: the model cannot see images. For tasks involving OCR, layout understanding, or scene recognition, it has no choice but to guess from context, and the results are often unreliable. dsh-visibridge was built for exactly this gap. It plugs into DeepSeek Harness as a host-level bundle, giving text-only models a reliable path to vision without modifying the model or the Harness core configuration.\n\nThe plugin follows an evidence-driven approach: the analyze_image tool sends an image to a vision backend (local Ollama, Xiaomi MiMo cloud, the official DeepSeek vision model, or any OpenAI-compatible endpoint) and retrieves a structured JSON evidence report covering full OCR text with line-by-line detail, layout blocks, scene and entity semantics, color and style notes, and self-reported uncertainty items. The text model then answers based on that evidence instead of fabricating. The capture_image tool goes further, letting the model autonomously snap a USB camera photo, analyze it, and compare across iterations, creating a modify-shoot-observe-refine visual loop ideal for screen debugging, document scanning, and any workflow that requires observing physical changes.\n\nIf you are building Agent workflows on DeepSeek Harness and want to give your text-only model eyes without introducing proxy processes or touching core configuration, dsh-visibridge offers a lightweight, zero-dependency host-level solution. The local Ollama backend runs with zero upload and auto keep-alive; cloud backends switch with a single config line; API keys are automatically redacted; and private-network endpoints are denied by default, balancing ease of use with security.