dsh-maclens

Image analysis on a Mac usually means picking one of two imperfect paths: call a cloud API, which is per-request priced, slow, and raises privacy concerns; or deploy a vision-language model, which is heavy, memory-hungry, and still not instant. dsh-maclens bridges the native Apple Vision framework directly into DeepSeek Harness so the model can invoke local OCR, image classification, face detection, and document layout parsing as simple tool calls. Every pixel stays on your machine; no API key, no network calls, no daemon.

The plugin exposes five tools: line-by-line OCR with confidence scores and normalized bounding boxes, image classification (document, chart, photo, and more), face-box detection with count, document parsing that recovers left and right column layout, and an all-in-one combined read. OCR ships with Simplified Chinese and English by default and supports over 30 recognition languages out of the box. For tall screenshots, a slice mode splits the image into overlapping strips, OCRs each one, and stitches results back to whole-image coordinates so small text survives the pipeline. All inference runs on the Apple Neural Engine, delivering sub-second results on typical screenshots with zero external dependencies.

If you need fast, private text extraction or face detection on macOS without uploading screenshots to a third-party server, dsh-maclens is built for that. For open-ended semantic understanding, such as narrating what a photo is about, pair it with a VLM bridge plugin for the best of both worlds.