Introduction

In a plugin ecosystem like DSH, a common pain point is that DeepSeek text models can read files but cannot directly view images. If error screenshots, UI drafts, and charts can only be described manually, the efficiency is low and prone to distortion. The approach of deepseek-vl-support is: instead of requiring the text model to natively view images, when it attempts to read an image, the plugin hands the image over to an optional vision service and returns the textual description back to the model.

What is this

limccn/deepseek-vl-support is an MIT-licensed plugin maintained by limccn, with the npm package name @limccn/deepseek-vl-support. It intercepts image read requests, sends them to the selected OpenAI-compatible vision endpoint, and then returns the result to the model in text format.

Core Features

  • Provides an image understanding path for DeepSeek text models: when the model reads an image, the plugin forwards the image to the selected vision service.
  • The content returned to the model is in a fixed format: [Vision of screenshot.png]: <description>.
  • Supports OpenAI-compatible vision endpoints, including OpenRouter, SiliconFlow, DashScope, Ollama, llama.cpp, VLLM, and LM Studio.
  • Supports multiple agents, including Claude Code, Codex, DeepSeek Harness, Cursor, GitHub Copilot, VS Code, OpenCode, and Trae.
  • Provides a one-time installation wizard, per-agent installation commands, and a doctor check command.
  • Comes with the deepseek-vision skill, using the standard Agent Skills layout.
  • Provides support for DeepSeek Harness via an in-process bundle patch.

Installation and Enablement

First, confirm your environment:

  • Node.js 18 or later.
  • A visual service account and API key. The installer will ask once and store it locally.

Run the installation wizard in your project directory:

cd path/to/your/project
npx @limccn/deepseek-vl-support@latest install

After installation, restart the session. Optional check:

npx @limccn/deepseek-vl-support@latest doctor    # look for [OK]

Installation per Agent

If you only need to enable it for a specific agent, you can use the corresponding command.

Claude Code:

npx @limccn/deepseek-vl-support@latest install --target claude

Codex:

npx @limccn/deepseek-vl-support@latest install --target codex

DeepSeek Harness:

dsh plugin --profile web add @limccn/deepseek-vl-support@latest

Other Installation Methods

If you only need the deepseek-vision skill, you can use the Agent Skills CLI:

npx skills add https://github.com/limccn/deepseek-vl-support

If the agent supports the Agent Plugins standard, you can also say in the conversation:

Install the plugin from https://github.com/limccn/deepseek-vl-support and enable it

After installing from GitHub, run the command once more to configure the visual endpoint:

npx @limccn/deepseek-vl-support@latest install --target <your agent>

Typical Usage

After installing and restarting the session, simply ask the model to read an image. The tool will hand the description returned by the vision service to the model in the following format:

[Vision of screenshot.png]: <description>

The model will then continue to answer based on this description.

Migration from 0.2.x to 0.3.0

0.3.0 moves the npm package to @limccn/deepseek-vl-support and publishes it to npmjs and GitHub Packages. CLI commands, configuration, cache, skill files, and tags remain unchanged; the 0.2.x package is kept frozen.

If you already have 0.2.x, you can run:

npx @limccn/deepseek-vl-support@latest install --update

Then check:

npx @limccn/deepseek-vl-support@latest doctor

Applicable Scenarios and Notes

Suitable for scenarios where DeepSeek text models are used in agents like DSH, Claude Code, Codex, etc., and need to understand image content such as error screenshots, UI drafts, charts, etc. Once configured, no additional maintenance of model settings or configuration files is required.

Suggestions before use:

  • Confirm that the selected visual service is available, and the API key is used only for local storage and calls.
  • When enabled in DSH, the plugin runs with the permissions of the current DSH process; check the source code and MIT license before installing.
  • Restart the corresponding session after installation or migration, then run doctor or verify by reading images.

Conclusion

The value of deepseek-vl-support lies in preserving the DeepSeek text model workflow while supplementing image understanding through an external vision service. The GitHub repository address is https://github.com/limccn/deepseek-vl-support; in the DSH community directory, it can be searched by the repository name limccn/deepseek-vl-support.