Introduction

The DSH plugin ecosystem emphasizes “everything is a plugin.” The community directory here is an independent site and is not officially affiliated with DeepSeek/Fangcun; this article introduces it solely based on the plugin facts.

For developers integrating ComfyUI into DeepSeek Harness (DSH), there are several common difficulties: the Agent needs to know which nodes the current server supports and what parameters the workflows have; generated results need to be previewed and downloaded in the panel; the API Key for remote ComfyUI should not be sent directly to the browser; and attention to detail for complex workflows is easily overlooked by the Agent.

dsh-comfyui encapsulates all these tasks into a single plugin: Agent tools, UI panels, workflow skill packages, and same-origin media proxy are all provided by the plugin.

What is this

dsh-comfyui is a DSH plugin maintained by fandc520, under the MIT license.

It enables the DeepSeek Harness Agent to intelligently drive local or remote ComfyUI to generate any content, complete with workflow management, asset management panels, and skill package mounting.

Below, we introduce its core capabilities, installation methods, typical usage, and precautions.

Versions, Dependencies, and Installation

First, confirm the version compatibility, then proceed with installation. The verified version requirements are as follows:

dsh-comfyui 0.4.0
DeepSeek Harness >= 0.1.2

dsh-comfyui 0.3.x beta line
DeepSeek Harness 0.1.1

If versions are incompatible, dshmarket will display a risk warning.

Installation command:

dsh plugin --profile web add dsh-comfyui

Runtime environment requirements include:

DeepSeek Harness web/desktop profile
Running ComfyUI server
Default server address http://127.0.0.1:8188
Node >=22.19
@deepseek-ai/cordis ^4.0.1
@deepseek-ai/dsh-settings ^0.1.2-alpha.4

If you want to use the video template, you also need:

ComfyUI-WanVideoWrapper
Wan 2.1 model

Agent Tools

The plugin provides four tools for the Agent.

comfyui_run

comfyui_run is used to submit generation tasks.

It can submit workflows in API format or use built-in templates:

txt2img
img2img
video

It supports two execution modes:

Synchronous wait
Asynchronous background task

comfyui_object_info

comfyui_object_info is used to list node definitions supported by the server.

The Agent can construct valid workflows on the fly based on node definitions, rather than guessing node names and parameter names out of thin air.

comfyui_workflow

comfyui_workflow is used to manage the workflow library.

It supports the following operations:

list
run
skill
refresh

comfyui_skill

comfyui_skill is used to read and write workflow skill packages.

It supports the following operations:

list
read
write
append
mkdir
rename
delete
enable
require

UI Panel

The plugin provides a UI panel with three tabs:

Workflow
Assets
Queue

The responsibilities of each tab are as follows:

  • Workflow: Workflow management.
  • Assets: Preview and download generated results.
  • Queue: Queue operations.

Loading Area

The loading area provides a media loader with visual selection options:

Image
Video
Audio

Added media assets will be automatically filled into the workflow’s loading parameters that are not explicitly specified, in order.

This way, the Agent does not need to guess filenames and can complete corresponding tasks based on the media in the loading area.

Workflow Skill Packages

Workflow skill packages are used to document the usage instructions for a specific workflow.

It supports the following capabilities:

On-demand three-level disclosure documentation
Panel editing
Agent writing
Pre-run mandatory reading validation

A parameter list can only indicate what parameters the workflow has, while a skill package can further supplement applicable scenarios, key parameters, and precautions. The Agent can also read and write skill packages via the comfyui_skill tool.

Media Proxy and Settings

Generated files are forwarded via a same-origin route:

/comfyui/media

It forwards based on filename and does not rely on ComfyUI’s in-memory history state.

A new ComfyUI section has been added to the settings page, allowing configuration of:

Server address
API Key environment variable name
Local ComfyUI directory
Media access address

If remote ComfyUI is located behind an authenticated proxy, the key can be provided via credential storage or an environment variable specified by apiKeyEnv. The key is never sent to the browser.

The API Key is not sent to the browser; the browser does not directly contact ComfyUI to avoid CORS and mixed content issues.

Typical Usage

You can directly tell the Agent the following requirements:

Use ComfyUI to draw a picture of a red cat

Turn this picture into a cyberpunk style

Turn this anime picture in the loading area into a realistic photo, with the same resolution as the original

Generate a 5-second short video: city at sunset

Run the Krea-Afterlight I previously saved in ComfyUI

These examples correspond to verified usage examples of the plugin, without making additional assumptions about extra operations.

Applicable Scenarios and Notes

Suitable for the following scenarios:

Have DeepSeek Harness web/desktop profile
Have a local or remote ComfyUI server
Want the Agent to directly submit txt2img / img2img / video tasks
Need to manage workflows, view assets, and operate the queue in the panel

Precautions:

  1. First, confirm the version pairing between dsh-comfyui and DeepSeek Harness.
  2. The video template depends on ComfyUI-WanVideoWrapper and the Wan 2.1 model.
  3. The runtime environment must meet Node >=22.19.
  4. The plugin runs with the permissions of the current DSH process; please check the source code and license before installation; the current license is MIT.
  5. If using remote authentication, the key is stored only on the server, in credential storage, or in environment variables, and is never sent to the browser.

Conclusion

The value of dsh-comfyui lies in transforming ComfyUI from a manual canvas operation into a generative backend callable by the DSH Agent: the Agent can submit tasks, view nodes, and manage workflows and skill packages; the panel can view workflows, assets, and the queue; media is accessed via a same-origin route, and the API Key is not sent to the browser.

Links:

GitHub: https://github.com/fandc520/dsh-comfyui
Directory Page: https://www.skillhub.cn/plugins/fandc520/dsh-comfyui

The directory page address comes from plugin clues; the data scraped this time does not directly contain this URL.