dsh-maclens

dsh-maclens

Bridge Apple's on-device Vision framework (macOS) into DeepSeek Harness (dsh) as local tools: OCR, image classification, face detection, and document layout.

• 🔒 100% local — no network, no API key, no daemon, no upload. The Vision framework is already on your Mac.
• ⚡ Fast — sub-second OCR for typical screenshots, running on the Neural Engine.
• 🇨🇳 Great Chinese OCR — Vision's recognition supports zh-Hans + 30+ languages out of the box.
• 🖼️ Tall-screenshot slicing — --slice splits very tall images (chat logs, scrolling captures) into overlapping strips so small text is not lost to Vision's downscale.

Tools

Tool	What it does
maclens_ocr	Transcribe all visible text with confidence + bounding boxes
maclens_classify	Top-N category identifiers (document, chart, diagram, photo, …)
maclens_faces	Face bounding boxes and confidence
maclens_document	OCR + simple left/right column layout analysis
maclens_describe	One call combining OCR, classification, faces and layout

Install

Prerequisites: macOS (Vision framework), Xcode Command Line Tools (Swift 6+).

# 1. install from npm (bin/maclens ships in the package — no build needed)
dsh plugin --profile desktop add dsh-maclens

# 2. or, from a checkout: build the Swift bridge yourself
cd dsh-maclens
bash scripts/build.sh          # → bin/maclens
dsh plugin --profile desktop add ./dsh-maclens

The plugin spawns bin/maclens (or swift/.build/release/MaclensBridge, or a maclens on PATH — set MACLENS_BIN to pin a specific binary).

Usage

Each tool takes an absolute image path. All OCR-based tools accept:

Argument	Meaning
languages	Comma-separated recognition languages (default zh-Hans,en-US)
maxLines	Cap OCR output (large screenshots)
slice	Slice tall images into overlapping strips for reliable OCR
sliceHeight	Strip height in pixels when slicing (default 4096)

maclens_describe additionally accepts top (number of classification categories).

How it works

dsh (text-only model)
  └─ maclens_ocr / classify / faces / document / describe   ← dsh tools (lib/index.js)
       └─ bin/maclens (Swift CLI)                           ← swift/Sources/MaclensBridge
            └─ Apple Vision framework (VNRecognizeTextRequest, VNClassifyImageRequest,
               VNDetectFaceRectanglesRequest)               ← on-device, offline

Positioning note: Vision is a computer-vision toolkit, not a vision-language model. maclens_* transcribes, classifies, and detects — it does not produce free-form semantic description ("what is this image about"). Pair it with a VLM bridge (e.g. modlens + qwen-vl) when you need open-ended understanding; use maclens_* when you want fast, free, private OCR and detection.

Development

cd swift && swift build -c release
.build/release/MaclensBridge ocr --image /path/to/img.png
.build/release/MaclensBridge classify --image /path/to/img.png --top 3
.build/release/MaclensBridge describe --image /path/to/img.png --slice

# tests
swift test --package-path swift

License

MIT