Introduction

In the DSH plugin ecosystem, a common practice is to encapsulate external capabilities into tools for models to call. Some DeepSeek models only process text; when encountering scenarios requiring “image recognition,” an additional bridging tool is required.

dsh-deepseek-vision is a DSH plugin: it registers a deepseek_vision tool, calls the local deepseek-vision-cli, and returns the image description as text to the model. It is important to clarify that this is not local recognition, but sending the image to the DeepSeek web endpoint; it relies on an unofficial web interface, so users need to be aware of risk control risks.

What is this

The repository name for dsh-deepseek-vision is Cheng-cheng9669/dsh-deepseek-vision, maintained by Cheng-cheng9669, and licensed under MIT.

The core problem it solves is: providing an available image recognition entry for DeepSeek models that can only process text. After the plugin registers the deepseek_vision tool, the model can submit local image paths and optional questions via this tool. The plugin then calls the local deepseek-vision-cli and finally returns the obtained image description text to the model.

Core Features

The following introduces its confirmed supported capabilities:

  • Registers the deepseek_vision tool.
  • deepseek_vision accepts local image paths and an optional prompt.
  • Supports PNG, JPG, WebP, GIF; does not support HEIC.
  • Calls the local deepseek-vision-cli, returning the image description as text to the model.
  • Supports installation from the local DSH plugin path or a release tarball.
  • Supports manual login helper and token login for DeepSeek web sessions.
  • Calls the external CLI via the DSH subprocess service, using an argv array, not the shell.

Installation and Usage

The documentation provides two installation commands, corresponding to local plugin directories and release tarballs, respectively.

Using a local directory:

dsh plugin --profile web add D:/Dsh/tools/dsh-deepseek-vision

Using a release tarball:

dsh plugin --profile web add dsh-deepseek-vision-0.0.1.tgz

Here only the commands appearing in the documentation are used, and no other installation addresses are concatenated based on the repository name.

It is recommended to check the source code, the behavior of external CLI dependencies, and the license before installation. The plugin license is MIT, but it calls an external CLI and runs with the current dsh process permissions; if the local environment is sensitive to permissions, you should first confirm the source is trustworthy and the behavior meets expectations.

First Login

DeepSeek web vision requires a web login session. The manual login helper command provided by the documentation is as follows:

py -3.13 D:\Dsh\tools\deepseek-vision-cli\dsv_manual_login.py

After completing manual login, the token is saved in the local path:

D:\Dsh\.cache\dsv_token

This token is only saved locally and will not be submitted to the repository.

If you already have an available DeepSeek web session, you can also use the token login method. The core prerequisite is still writing the valid token to the local token file.

Typical Usage

The tool signature after the plugin is registered is as follows:

deepseek_vision(image, prompt?)

Parameter description:

  • image: Absolute path of local image, supports PNG/JPG/WebP/GIF, does not support HEIC.
  • prompt: Optional, used to specify a question or analysis requirement for this image.

After calling, the plugin returns the obtained image description as text to the model. That is to say, the model ultimately receives a text result, not the image itself.

Suitable Scenarios and Notes

Suitable users for this plugin include:

  • Already using DSH and wishing to integrate external CLI capabilities into the model toolchain.
  • The current DeepSeek model only processes text and needs to supplement image recognition capabilities.
  • Can accept the risk control risks brought by the unofficial web interface and can control the calling frequency.

Notes to pay special attention to:

  • Images are sent to the DeepSeek web endpoint; this is not local recognition.
  • This capability relies on an unofficial web interface and may trigger risk control; it is recommended to use a secondary account and maintain low-frequency usage.
  • HEIC format is not supported.
  • The plugin calls the external CLI via the DSH subprocess service, using the argv array, not the shell.
  • The repository description states: static security audit reports critical errors for starting subprocess; the code has been manually reviewed, only running fixed dsv.py, with no command concatenation or extra data exfiltration.
  • The plugin runs with the current dsh process permissions; source code, external dependencies, and the license should be checked before installation.
  • The DSH plugin ecosystem emphasizes “everything is a plugin,” but the third-party plugin directory is not equivalent to the DeepSeek / Fangyuan official app store.

Links

GitHub Repository:

https://github.com/Cheng-cheng9669/dsh-deepseek-vision

Plugin directory clue address:

https://www.skillhub.cn/plugins/Cheng-cheng9669/dsh-deepseek-vision

The directory page address comes from the plugin clue and was not directly confirmed in the already captured data; please verify it yourself before accessing.