dsh-local-ocr

When you ask a DeepSeek Harness agent to read text out of a local screenshot or a scanned document image, the most natural expectation is that it should simply recognize the characters. dsh-local-ocr exists for exactly that: it wraps the operating system built-in OCR into a single model-facing tool, so the agent can pass an absolute image path and receive the recognized text along with a confidence score, with no cloud round-trip involved.

Under the hood the plugin bridges macOS Vision OCR and Windows Media OCR through @napi-rs/system-ocr, exposing one unified Node.js API. The single tool it registers, local_ocr_image, uses a lazy-load strategy: the native add-on is loaded only the instant the model actually invokes the tool, keeping idle overhead at zero. Return fields are deliberately minimal and stable, namely text and confidence. An image with no detectable text yields an empty string and a confidence of 0 rather than throwing an error that could derail a conversation.

This is a good fit for developers running DeepSeek Harness on macOS or Windows who want image text recognition to stay entirely on-device. The plugin ships as a self-contained bundle that attaches to a Harness profile without modifying the Harness source, without introducing a cache layer, and without deciding whether OCR should run, leaving timing entirely to the agent. Linux is not yet supported in this MVP; if you just need a lightweight, locally-runnable OCR entry point that slots seamlessly into your existing Harness workflow, it is precisely enough.