dsh-mac-vision

Personal / macOS AI Plugin

A native macOS vision plugin that gives text-only DeepSeek Harness models local OCR and visual evidence.

In brief

A DeepSeek Harness plugin for inspecting local images, clipboard content, screens, and application windows through on-device Apple Vision.

Built on Apple Vision so source images do not need to be uploaded to a vision-model service.

Text-only agent models can reason about local workflows but cannot directly inspect a screenshot, clipboard image, or application window. Sending that material to a separate vision-model service also introduces an avoidable privacy boundary.

dsh-mac-vision adds a native visual evidence layer to DeepSeek Harness. One plugin installs the capture and Apple Vision tools together with model-facing guidance for deciding when to use them and how to communicate uncertain results.

A useful local vision tool has to coordinate several concerns at once: macOS capture permissions, native OCR, small-text recovery, structured tool output, cancellation and timeout behavior, and careful separation between direct evidence and model interpretation.

What it includes

  1. 01On-device OCR for files, clipboard images, screens, and selected windows
  2. 02Structured evidence with confidence, coordinates, alternatives, and review state
  3. 03Automatic crop-and-upscale refinement for small text and notation
  4. 04Bundled Skill for separating observed, inferred, and uncertain results

How it works

  1. 01Select a local source

    Accept an image file, clipboard image, full screen, frontmost application, or selected visible window.

  2. 02Capture on macOS

    Resolve the requested source locally and respect the operating system's Screen Recording permission boundary where required.

  3. 03Run native analysis

    Use Apple Vision for OCR and optional classification, barcode, saliency, people, or animal detection.

  4. 04Refine difficult regions

    Crop and upscale small text or notation, then run a second OCR pass while retaining conflicting candidates.

  5. 05Return evidence to the model

    Provide structured observations and guide the agent to distinguish direct evidence, interpretation, and uncertainty.

Build notes

Note 01

Keep image analysis on device

local by default
Context
Screenshots and local files may contain private workspace, account, or application data.
Approach
Use the macOS Vision framework for OCR and supported detections instead of uploading source images to a vision-model service.
What it enables
The visual source remains on the Mac while the agent receives structured evidence it can reason over.
Note 02

Bundle execution tools and usage policy

agent integration
Context
A native tool can return evidence, but a model still needs guidance about when to call it and how to report ambiguous results.
Approach
Ship the vision tools and their model-facing Skill as one installable Harness plugin.
What it enables
Users install one package, while the runtime loads both execution capability and consistent evidence-reporting behavior.
Note 03

Preserve uncertainty in structured output

evidence design
Context
Small text, formulas, and visually ambiguous regions can produce plausible but conflicting OCR candidates.
Approach
Return confidence, coordinates, alternatives, refinement state, and explicit distinctions between completed and unrun detectors.
What it enables
The agent can identify what was directly observed, what was inferred, and what still requires review.

Compatibility

  • The plugin requires macOS, Node.js 22 or later, DeepSeek Harness, and a local Swift compiler.
  • Screen and application-window sources may require Screen Recording permission; local image files do not.
  • Small text, formulas, and ambiguous regions can still require a targeted second inspection and human review.
Next project13dsh-trading212Personal / Fintech Workspace

Project image

Interactive demo