dsh-macos-vision-ocr

Reading text out of screenshots, scans, or document images on a Mac typically means shipping pixels to a cloud OCR service, which is an uncomfortable trade-off when the content is confidential. dsh-macos-vision-ocr bridges Apple's macOS Vision framework into DeepSeek Harness, giving every text model an ocr_image tool that runs the entire pipeline on-device: no API key, no outbound requests.

The plugin invokes VNRecognizeTextRequest at its highest-accuracy preset and handles PNG, JPEG, WebP, GIF, TIFF, BMP, HEIC, and HEIF inputs. Callers can set BCP-47 recognition languages on each invocation, so a single profile can process Chinese, English, Japanese, or any language the Vision stack supports. A compact Swift helper is compiled once and cached by content address for instant reuse. Every call returns a bounded text payload with a truncation flag, and the subprocess is always spawned with a fixed argument vector—image paths never pass through a shell interpreter.

It is designed for macOS 13+ users who need reliable, privacy-preserving text extraction as a tool inside an agent workflow: contract review, receipt digitization, multilingual document triage. It is not a visual-understanding engine—it will not detect objects, faces, or scene semantics, and reading order is a geometric approximation that can falter on multi-column or heavily stylised layouts. If your priority is keeping pixels and text on your own hardware while still feeding structured text to a model, this plugin is a focused, MIT-licensed way to get there.