dsh-mac-vision

The biggest gap in pure-text models is that they cannot see. When you want the model to read a screenshot, locate an error message on screen, or extract formulas from a whiteboard photo, you are typically forced into third-party OCR or multimodal APIs, incurring per-call fees while image data leaves your machine. dsh-mac-vision gives any pure-text model running on macOS a native, fully local vision capability and closes that gap in a single install.

Built on Apple's Vision framework, the plugin reads local images, clipboard contents, the full screen, or a specific application window, performs OCR, and returns structured evidence including text, confidence scores, bounding-box coordinates, and re-check status. When it encounters small-font or formula-dense regions, it automatically crops and upscales the area for a second-pass recognition. Optional detectors cover image classification, bar-code reading, salient-region detection, and person or pet detection, so the model can pull multi-dimensional visual cues in a single call. Every result is explicitly tagged as directly observed, semantically inferred, or uncertain, preventing the model from hallucinating details it never actually saw.

If you are already using DeepSeek Harness on a Mac and regularly need screen analysis, document OCR, debugging-screenshot interpretation, or whiteboard transcription, this plugin is the missing piece. It is completely free under the MIT license, requires no vision API key, downloads no third-party models, and keeps all image data on your machine. Install it once and you are ready to go with no monthly subscription, no per-call charges, and no extra configuration.