dsh-omni-vision: Local Vision Capabilities for Pure-Text Models
The `dsh-omni-vision` plugin for DeepSeek Harness (DSH) provides local visual perception capabilities for pure text models and supports the Windows platform. It includes four core tools: `eyes_render` for generating images through drawing; `eyes_paste` for pasting images; `eyes_ocr` for recognizing text in images using Windows engines; and `eyes_analyze` for parsing pixel data. It supports scenarios such as OCR recognition, text transcription, flowchart analysis, and pixel layout analysis. Before use, it must be installed via command and the service restarted. It relies on a browser GUI and supports only 8-bit PNG format.
Read MoreDSH Image-to-Text Input Plugin: Intercept Image Input for Text-Only LLMs
The dsh-plugin-image-input plugin converts images into structured textual descriptions by locally invoking a vision API, enabling pure text large models (such as DeepSeek R1) to handle image input. Users only need to paste or drag and drop images; the plugin can automatically or manually convert them into textual context sent to the model, with support for multiple images. After installation, an OpenAI-compatible vision API must be configured. The plugin features multimodal compatibility, fault tolerance, and safety restrictions, effectively resolving request failures caused by the model's inability to recognize images.
Read Moredsh-design-qa: Enabling Image Understanding for Text-Only Models in DeepSeek Harness
The dsh-design-qa plugin enables text-only models to “read” images by integrating multimodal tools, allowing them to perform visual judgments within the product design loop. It supports on-demand invocation, does not consume the main model's context, and includes built-in evaluation benchmarks. Installation requires configuring an Alibaba Cloud Bailian API Key, and it defaults to using the qwen3.8-max model. The plugin emphasizes that “visibility does not equal judgment capability” and should be used according to specific scenarios.
Read MoreUse dsh-drop-to-path to convert images and files to paths for sending to pure text models
`dsh-drop-to-path` is a DeepSeek Harness Web plugin (MIT licensed) maintained by loudMore. When a plain-text model attempts to send images, they will be blocked by the host's pre-check, and files such as PDFs cannot be sent as attachments either. This plugin retains the native attachment bar, and when sending, it writes images and files to the `.drops/` workspace directory and attaches their paths. The directory installation command is `dsh plugin add github:loudMore/dsh-drop-to-path`; it is recommended to use it with `dsh-vision-toolkit`, and please read `ADAPTING.md` before upgrading DSH.
Read MoreEquip Plain-text Coding Agents with a Visual Toolbox Using agent-vision-toolkit
Agent-Vision-Toolkit is an open-source visual toolkit (MIT License, Python) maintained by Anionex. It provides skills such as glance/ground/detect/trace/crop and vision-tools for pure text-coding intelligent agents like DeepSeek, covering visual question answering, long-screenshot OCR, UI restoration and GUI operation, and can be connected to Codex, Claude Code and other models. The community directory installation command points to this repository; the dsh Web/Headless native Bundle requires additional installation of @anionex/dsh-vision-toolkit.
Read More