dsh-tool-see-image：为纯文本 DSH 会话补上「看图」能力
DSH 默认纯文本模型无法直接处理图像。dsh-tool-see-image 在 Host 侧注册 see_image：通过 ctx.fs 读取沙箱内图片，调用 OpenAI 兼容视觉接口（默认智谱 GLM-4V-Flash），将文字描述回传给会话。本文说明工作原理、web profile 下的复制安装、cordis.patch.yml 配置项、典型调用与卸载步骤，并提醒 API 密钥与进程权限、单图 15MB 与 maxTokens 限制。适合不想换多模态主模型、仅需按需识图的 DSH 用户。
Read MoreClaw Eyes 虾看：让文字模型通过剪贴板「看图」
Claw Eyes 虾看（@user_bed45be1/claw-eyes，v2.0.1）是 SkillHub 社区 ai-agent 类 Skill，在 Windows/macOS/Linux 读取剪贴板图片并优先通过视觉 API 分析，MCP 备用。支持中英触发如「看图」、多供应商预置与自定义提示词，需配置 API Key。文章介绍能力边界、skillhub CLI 安装到各 Agent skills 目录的方法，以及典型截图分析流程与使用注意。
Read Moredsh-computer-use-native: A Vision-Based Windows Desktop Control Plugin
`dsh-computer-use-native` is a Windows desktop automation plugin designed specifically for DeepSeek Harness (DSH). To address the issue that modern applications (such as Chromium and Electron) do not expose accessibility trees, the plugin captures real pixels via screenshots, drives windows through Win32 FFI, and provides core tools for screenshotting, clicking, and input. It supports background input mode to avoid stealing focus, includes viewport coordinate mapping to ensure precise model operations, and optimizes screenshot performance using pixel differential comparison, significantly reducing interaction costs. This solution is compatible with Windows 10/11 and Node.js 22+, offering an effective way for visual models to control modern Windows applications.
Read Moredsh-tool-see-image: Adding Vision Capabilities to DeepSeek Harness
`dsh-tool-see-image` is a plugin tool for DeepSeek Harness that addresses the limitation of pure text models in directly processing images. Acting as an intermediary layer, it invokes a visual model (GLM-4V-Flash by default) to analyze images and return text descriptions. Core features include reading files through the `ctx.fs` sandbox, supporting OpenAI-compatible visual model interfaces, and providing an additional paste-to-text script. Installation requires manually copying the plugin files, declaring dependencies, configuring `cordis.patch.yml`, and setting an API key. Configuration options support customizing the `baseURL`, model, and token limit. Released under the MIT License, it aims to enhance model visual capabilities through lightweight extension.
Read MoreDeepSeek Harness Vision Bridge Plugin: Routing Image Messages to OpenAI-Compatible Endpoints
The `zh-u-hb/dsh-vision-bridge` plugin for DeepSeek Harness (DSH) listens to the DSH message pipeline to address the issue that text-only models cannot process images directly. When it detects that the active model does not support images, the plugin forwards the image to a user-configured OpenAI-compatible vision model and substitutes the returned text back into the conversation flow. It allows users to configure the API endpoint, API key, and model ID via a web page, and automatically assembles prompt and image data. Note that it supports only the OpenAI protocol and handles image messages only at specific pipeline stages.
Read Moredsh-image-bridge: Converting Images into Text Descriptions So Pure-Text Models Can Also “See” Images
`dsh-image-bridge` is a DeepSeek Harness plugin designed to address the inability of pure text models to understand images in Web GUIs. It intercepts message hooks, uses a vision model to transcribe images into textual descriptions, and replaces the original image block in the conversation. The plugin implements resilience strategies including automatic retries, model fallback, result caching, and graceful degradation on failure, and supports configuration of the primary model, fallback chain, and payload resolution limits. Once installed, it enables pure text models, such as DeepSeek, to “see” images.
Read Moredsh-view-image: Enabling Text-Only Models to Read Images
`dsh-view-image` is a DeepSeek Harness (DSH) plugin designed to enable text-only main conversation models to handle images. By configuring a separate OpenAI-compatible vision model, the plugin converts image data into plain-text descriptions and inserts them into the conversation history, giving text-only models the ability to “read” images. It supports paste-and-use in the Web UI and invocation via file paths in the terminal, and provides thumbnail preview, one-click copy, and persistent storage. Users need to install the plugin and configure a vision model endpoint; images are stored using content-addressed storage, but the attachment directory must be cleaned up manually.
Read Moredsh-vision-bridge: Enabling Text-Only Models to Receive Images
`dsh-vision-bridge` is a DeepSeek Harness plugin designed to address the inability of text-only models to process images. It leverages the `qwen2.5vl` model run locally via Ollama, converting images into text descriptions sent to the DeepSeek API to enable localized visual processing. The plugin provides a native image experience, transparent request-layer adaptation, structured output, and coordinate support. It includes nine vision tools, such as description, OCR, and region queries, and supports caching of duplicate images. Once installed, the plugin registers a “twin” route and injects a tool set, with all operations completed locally to ensure privacy and security.
Read Moredsh-image-vision: Integrating Vision Capabilities into Text-Only Models
`dsh-image-vision` is a lightweight plugin for DeepSeek Harness. By wrapping the `llm` service interface, it enables text-only primary models to understand images without requiring any modifications to host code or disrupting the official plugin ecosystem. It supports general image input, Feishu document parsing, and local image description, relying on visual model routing configurations with a `VL` suffix. Easy to install and highly compatible, it is ideal for integrating vision capabilities into agents without compromising the existing architecture.
Read MoreDeepSeek Harness Plugin dsh-subagent-vision: Enabling Pure-Text Main Agents to Read Images Directly in Conversations
The DeepSeek Harness plugin `dsh-subagent-vision` addresses the inability of pure-text models to handle images. It delegates image processing tasks to a configured vision subagent, enabling the main agent to remain text-only while receiving visual results. The plugin supports pasting, dragging and dropping, and importing images via file path; provides a native Web GUI experience; and enforces a 25 MB image size limit. After installation, users only need to describe the task to the agent, and the plugin automatically invokes the subagent's `read_image` tool to complete it.
Read MoreDuhai Vision: Universal Vision Model Adapter for DeepSeek Harness
`duhai-vision` is a general-purpose visual model plugin compatible with Codex and DSH, designed to provide auditable visual processing for long documents, tables, and complex layouts. It prioritizes PaddleOCR-VL by default, uses Qwen as a fallback, and supports returning structured data. Installation requires configuring a Paddle Token and adding the plugin via the command line. The image processing workflow is: save the image hash path -> extract structure with PaddleOCR-VL -> reason with DeepSeek. The plugin offers a free quota of 3,000 pages, but users should be aware of privacy risks associated with remote services.
Read Moredsh-image-bridge: Enabling Image Input Support for Text Models in DeepSeek Harness
`dsh-image-bridge` is a DeepSeek Harness (DSH) plugin designed to address the limitation that text-only models in DSH cannot process images. It wraps interfaces and leverages DSH’s surface replace mechanism to allow text models to receive images, save them to disk, and convert them into a text format for the model to invoke vision or MCP tools. The plugin supports image thumbnail display and is easy to install, but users should note limitations when mixing commands with images and when handling previous images.
Read Moredsh-tool-vision-read: A DSH plugin that enables text-only agents to "see" images
# Preface In agent development, a common gap arises: the model router for the main agent does not declare image inputs and is a pure text model. When users provide an image path for description, the agent cannot handle it—requiring manual switching to a session with vision capabilities or the agent manually viewing the image and relaying the information. The philosophy behind DeepSeek Harness (hereinafter referred to as DSH) is "everything is a plugin." This specific capability gap is ideally suited to be patched by a small plugin. Below, we introduce `dsh-tool-vision-read`, maintained by Mappedinfo. It registers a `vision_read` tool to route the image reading step to a dedicated vision model, enabling a pure text agent to obtain a textual description of the image with a single call.
Read Moredsh-image-vision-bridge: Converts images to text descriptions for the DSH main model
# Introduction In agent development with DeepSeek Harness (DSH), a common frustration arises: the primary model is a pure text model, yet users upload screenshots to the conversation. DSH 0.1.1+ implements a built-in fallback for this scenario—requests do not error out, but the model can only see an image placeholder and cannot access the actual image content. To enable the model to truly respond based on images, one must either switch to a multimodal main model or implement a custom forwarding logic layer. dsh-image-vision-bridge solves this problem: it automatically forwards images from user messages to a vision model (defaulting to mimo-v2.5 on the opencode-go route), feeds the returned text description to the main text model, and preserves the original image display in the chat history. DSH's philosophy is "Everything...
Read Moredsh-plugin-mm-vision: Adding Structured Vision Capabilities to Text-Only LLMs
dsh-plugin-mm-vision is a DeepSeek Harness (DSH) plugin designed to provide image understanding capabilities for pure text LLMs. The plugin registers the `mm_vision` tool, invoking an external vision model to convert images into structured spatial textual descriptions containing precise percentage coordinates, enabling text models to analyze spatial relationships without directly processing the images. Core features include "synesthesia encoding," applicable to candlestick charts, UI screenshots, and natural photos; it supports pixel-level coordinate localization and an optional 40x30 pixel color block grid reconstruction, and features an automatic mode that adapts output details based on the characteristics of charts or natural images. The plugin supports configuring any OpenAI-compatible vision model and provides a TTL cache mechanism to improve response speed. Installation requires the DSH environment and Node.js.
Read Moredsh-vision-provider: Selecting Vision Model Combinations in the DeepSeek Harness Web UI
dsh-vision-provider is a community plugin for DeepSeek Harness (DSH), designed to enable a dual-model bridge of "vision model analysis + DeepSeek reasoning responses." Rather than native pixel input, it exposes third-party vision models declared with image input capability in Settings as an optional combination. When a user sends an image-containing request, the selected vision model processes the image, and the result replaces the original image before being handed over to DeepSeek for reasoning, tool invocation, and final responses; plain-text requests are handled directly by DeepSeek. The plugin supports configuring third-party vision models via the Web UI or directly connecting to OpenAI-compatible endpoints. Installation requires Node.js >=22.19.0 and a compatible DSH version. Usage:
Read Moredsh-vision-opencode: Configuring Switchable Vision Models for Text-Primary Master Models
dsh-vision-opencode is a DSH Web plugin that converts chat images into text via configurable vision models, enabling pure-text main models like DeepSeek to continue reasoning without switching the primary model. It supports image recognition model selection in the input box, vision settings, the vision_read_image tool, and vision-image-analysis skill, including a 60-second timeout and failure retry fallback. Installation: dsh plugin --profile web add -w github:poiuyjie/dsh-vision-opencode. MIT license, maintained by poiuyjie.
Read MoreAdd Image Recognition Capability to Plain-Text DeepSeek Harness Using dsh-vision
Maintained by william-jin-cmu, dsh-vision adds a `view_image` tool for plain-text DeepSeek, forwarding local images or URLs to any OpenAI-compatible VLM and returning answers in text. The default free tier of Zhipu AI, glm-4.6v-flash, can be replaced with Bailian, Fangzhou or local Ollama. The community directory installation command is `dsh plugin add github:william-jin-cmu/dsh-vision`. The plugin runs with the permissions of the current dsh process, cloud paths will send images to third-party endpoints, so please check the source code and license before installation.
Read MoreEnable Image Recognition for Plain Text Conversations in DeepSeek Harness with dsh-vision-proxy
DSH-Vision-Proxy is an MIT-licensed DeepSeek Harness interface enhancement plugin maintained by Flyvhidbwo. It registers the `deepseek-vision` route, converts GUI-attached images into text via an OpenAI-compatible VLM, and then submits the content to the plain-text DeepSeek model for response. By default, it uses Bailian Qwen3.7-Flash, and can automatically detect local Ollama when no API key is available. This article verifies the installation commands, configuration methods and privacy boundaries against the community directory page, GitHub README and package.json.
Read More