DeepSeek Harness (DSH) decomposes agent capabilities into pluggable components, but many developers still rely on pure text models in their daily work: they cannot directly paste images in conversations, have to manually type file paths for error screenshots, must OCR long screenshots containing chat histories, and have no way to restore frontend pages from design drafts. In the community directory, there is a plugin with over 800 stars—dsh-vision-toolkit (maintained by Anionex)—with a clear objective: to equip text models in DSH with “eyes” and integrate capabilities such as image viewing, localization, cropping, and pixel comparison into a composable toolkit.

It should be noted that DSH’s core philosophy is “everything is a plugin,” and community directories like the SkillHub Plugin Library are maintained by third parties, with no official affiliation to DeepSeek or High-Flyer. Before installation, it is recommended to review the source code and MIT license yourself, and remember that plugins run with the permissions of the current dsh process.

What Is This¶

dsh-vision-toolkit is a native integration for DeepSeek Harness derived from the upstream project agent-vision-toolkit. It does not simply pass images to multimodal models for generic descriptions; instead, it extracts evidence tailored to the current task—such as “where is the error?” or “what are the button coordinates?”—and combines local image processing with the vision-skills Skill to let the Agent know when to view images, when to localize elements, and when to perform pixel diffs.

The project offers two layers of capabilities:

10 visual tools + vision-skills Skill: Covering playbooks for Q&A, OCR, UI restoration, GUI operations, and more.
Native DSH integration: Interfacing with Profiles, Sessions, Settings, Artifacts, and Web UI; pasting images can automatically route to the (Vision Toolkit) variant (transparent variant routing is enabled by default from 2026-08-19, and the model list can retain only the original name).

In the SkillHub directory, this plugin is categorized under “Model Inference”; the GitHub repository currently has approximately 821 stars and 34 forks, the npm package version is 0.1.38, and the license is MIT.

Core Features and Highlights¶

Paste Images, Ask Directly¶

In DSH Web, pasting a screenshot allows a pure text model to automatically switch to the corresponding Vision Toolkit variant, retaining thumbnails, conversation history, and workspace paths, without needing to manually copy file paths or switch models.

Task-Oriented Viewing, Not Generic Descriptions¶

Most visual bridges have multimodal models generate generic descriptions before passing them to text models, introducing an extra layer of semantic loss. Vision Toolkit passes the user’s message or the Agent’s call reason as a focus hint to the visual model, outputting key information relevant to the current step, resulting in fewer tokens and more targeted responses.

Ten Tools, Usable Individually or in Combination¶

Tool	Typical Question	Main Result
`vision_glance`	What’s happening in the image?	Targeted answer, OCR, multi-image comparison
`vision_ground`	Where is the target?	Original image pixel coordinates, optional bounding box preview
`vision_detect`	What buttons/icons are there?	Numbered element list with coordinates
`vision_crop`	Crop a specific region	PNG/JPEG cropped image
`vision_trace`	Convert graphics to vector	SVG
`vision_pixel_diff`	What’s the difference between implementation and reference?	Difference ratio, heatmap, JSON
`vision_long_screenshot_ocr`	Read a long screenshot	Markdown, chunked and audit output
`vision_extract_foreground`	Extract the main subject	Transparent PNG
`vision_dominant_colors`	What are the main colors?	Color palette or ranked candidate colors
`vision_html_screenshot`	Render a local page screenshot	PNG, optional `pageHeight`

Coordinates are uniformly given as original image pixels x1,y1,x2,y2, which can be directly passed to cropping or subsequent automation.

vision-skills: A Methodological Operations Manual¶

The Skill bundled with the plugin carries upstream playbooks, such as: long screenshot OCR, screenshot-to-UI restoration, icon/logo vectorization, sketch-to-structured code (Mermaid/Graphviz), screenshot-based GUI operations, and more. Adding “use vision-skills” to the prompt guides the Agent to select tools step-by-step and verify results.

Typical Scenario Demonstrations (from Official README)¶

Screenshot → Editable HTML: Left is the reference page, right is the HTML/CSS restoration, which can undergo further screenshot rendering + pixel comparison iteration.
Hand-drawn Draft → Usable Interface: Sketches directly become interactive frontend pages.
Rapid UI Restoration: Produces a first-version screenshot with similar layout in about three minutes (allowing approximate colors and icon libraries).

Installation and Enabling¶

The official README recommends installing via npm package name to a specified Profile (replace web/headless/desktop etc. with your actual profile name):

dsh plugin --profile web add @anionex/dsh-vision-toolkit

Headless Profile:

dsh plugin --profile headless add @anionex/dsh-vision-toolkit

DSH Desktop users need to open the DSH Terminal from the tray and execute the command in the built-in CLI of the desktop version (the desktop version does not write to the system PATH):

dsh plugin --profile desktop add @anionex/dsh-vision-toolkit

The DeepSeek Harness Plugin Library directory page provides another method: installing from GitHub:

dsh plugin add github:Anionex/dsh-vision-toolkit

For reproducible installations, you can pin the commit hash after the command, e.g., dsh plugin add github:Anionex/dsh-vision-toolkit#commit.

After installation, restart the corresponding Profile, open Settings → Vision Tools, configure the vision model provider (API Key stored as DSH Credential), and run Test Vision Model to confirm connectivity. On first launch, it prepares an isolated Python runtime: preferring system Python 3.11+, otherwise downloading a ~35MB managed Python from domestic mirrors (falling back to GitHub release sources if that fails), and installing locked dependencies like Pillow, NumPy, vtracer, etc.

Typical Usage Examples¶

In a conversation, paste a screenshot or place an image in the workspace and call /vision-skills, then describe the goal in natural language:

Look at this screenshot and tell me the cause of the error and what's worth fixing first.
Find the login button in the top right corner, return its original image pixel coordinates, and generate a preview image with a bounding box.
Crop this icon and convert it to SVG.
Restore the page based on reference.png, take screenshots after each round and perform pixel comparison until major differences disappear.

An example of Skill-guided UI restoration:

(Use vision-skills) Restore this image into HTML.

You can also configure the vision provider (OpenAI Chat Completions compatible or Anthropic Messages) in a Profile patch:

- id: vision-toolkit
  config:
    provider:
      baseUrl: https://api.example.com/v1
      credential: MY_VISION_KEY
      model: your-vision-model
      protocol: openai

The official documentation provides a step-by-step tutorial with images for obtaining an AIHubMix API Key and configuring Gemini 3.7 Flash for vision.

Applicable Scenarios and Precautions¶

Who It’s For:

Those who use pure text models daily for coding and debugging but want a “paste-and-ask” experience similar to multimodal models.
Those needing to extract chat/document content from long screenshots, restore UI from design drafts or screenshots, perform icon vectorization, GUI screenshot operations, and other higher-value visual tasks.
Harness users who want to use the same visual toolchain across DSH Web / Headless / Desktop.

Precautions:

Plugins run with the current dsh process permissions. Please read the source code and license before installation; do not blindly install unvetted repositories.
Visual capabilities depend on your configured multimodal endpoint; each check sends only necessary intent and images, with no context accumulation between calls, so extra costs are usually controllable. You can also use a local small multimodal side model to further reduce costs.
HTML page screenshots require local Chrome/Chromium/Edge; if missing, only this feature is unavailable, while other tools remain usable.
For local OpenAI-compatible services (LM Studio, Ollama, etc.), the baseUrl must include the /v1 prefix, otherwise model tests may return incompatible response structure.
DSH Desktop 2.0.1’s built-in plugin marketplace has known installation issues; it is recommended to install via terminal commands first.

Summary¶

If you are already using DeepSeek Harness but often find yourself stuck because “the model cannot see images,” dsh-vision-toolkit is worth shortlisting: a one-line command to integrate, paste images to ask questions directly, ten tools covering the full chain from OCR to UI restoration, plus vision-skills to solidify experience into an executable operations manual for the Agent.

SkillHub Directory Page: https://www.skillhub.cn/plugins/Anionex/dsh-vision-toolkit
GitHub Repository: https://github.com/Anionex/dsh-vision-toolkit
Upstream Toolkit: https://github.com/Anionex/agent-vision-toolkit
Project Website: https://agent-vision.anionex.me

Paste Images and View: Getting Started with the DeepSeek Harness Visual Plugin dsh-vision-toolkit