dsh-macos-vision-ocr

Working with screenshots, scanned pages, or document images on macOS usually means routing them through a cloud OCR API, which pulls sensitive content off your machine. dsh-macos-vision-ocr wires Apple's Vision framework directly into DeepSeek Harness so any text model can extract readable text from an image entirely offline—no API key, no network round-trip.

Under the hood the plugin runs VNRecognizeTextRequest at its most accurate setting and accepts a wide range of formats including PNG, JPEG, WebP, GIF, TIFF, BMP, HEIC, and HEIF. Recognition languages are specified per call using BCP-47 tags. On first use a small embedded Swift helper is compiled and cached by content address; subsequent calls reuse that binary. Output is capped in size with a truncation flag, and the subprocess is always launched with a fixed argument vector so image paths are never interpolated into a shell string.

If you regularly work with contracts, paper-take-home assignments, or multilingual receipts on macOS 13 or later and would rather keep those files local, this plugin fills the gap. It is purpose-built for pixel-to-text extraction—not object detection, scene understanding, or layout interpretation—and respects Harness filesystem policies throughout. A natural fit for privacy-conscious developers and analysts who need OCR as a building block rather than a standalone product.