dsh-windows-ocr

When using DeepSeek Harness, many text-only models cannot directly process image attachments, forcing users to rely on vision-capable models or manually convert images. The dsh-windows-ocr plugin solves this problem: it leverages the built-in Windows OCR engine to recognize text from images locally on your device, sending only the extracted text to the model. This allows any text-only model to 'understand' image content while ensuring that raw image data never leaves your machine.

The plugin's core strength lies in its privacy-first local processing. By default, all images are OCR'd locally, with no image bytes sent to model providers, making it ideal for handling sensitive documents or private data. It integrates seamlessly into the dsh workflow, automatically intercepting image attachments and replacing them with readable text without requiring changes to model configurations. Additionally, it supports multiple languages (via Windows language packs) and offers an opt-in passthrough mode for genuine vision models.

This plugin is particularly suited for DeepSeek Harness users who prioritize privacy and need to process images locally. Whether you're an analyst wanting text models to analyze charts, or a developer building secure workflows, it extends your capabilities without sacrificing convenience. If you use text-only models and frequently handle images, dsh-windows-ocr provides a reliable way to bridge that gap while keeping all data under your control.