dsh-umi-ocr-vision

In DeepSeek Harness, if your primary model is text-only (say deepseek-official), screenshots and document images you drop into a chat simply get ignored. dsh-umi-ocr-vision was built for exactly this gap: borrowing the bridging idea from dsh-vision, it detects that the current model cannot handle images, then hands them to a locally running Umi-OCR instance for text extraction. The OCR output is injected as untrusted visual context into the request, and your original text model produces the answer. Everything stays on your machine — HTTP mode defaults to 127.0.0.1, and no image ever leaves your host.

The plugin never replaces the model you picked in the Harness UI, keeping the behavior transparent. Multiple chat attachments are OCR’d in order and merged into a single request. OCR results are explicitly tagged as untrusted observational data, so any prompt text baked into an image cannot gain system privileges. Beyond auto-bridging, enabling enableVisionTools exposes a full toolkit — vision_ocr, vision_glance, vision_detect_text, vision_ground_text, vision_crop, vision_long_screenshot_ocr, vision_dominant_colors, vision_pixel_diff — covering everything from single-image recognition to chunked long-screenshot OCR and pixel-level diffing, all offline.

It is best suited for users whose daily stack is a text-only DeepSeek model, who regularly need to read screenshots, tables, CAPTCHAs, and UI text, who care about privacy and want images to never leave localhost, and who do not want to switch to a heavyweight multimodal model just for an OCR pass. If you need to understand spatial relationships or object semantics in an image, reach for a true vision-model plugin like dsh-vision instead.