Introduction

DeepSeek Harness (DSH) extends its functionality through a plugin mechanism. dsh-jimeng is a plugin that integrates Volcengine Ark’s Jimeng (Seedream) image generation model into DSH. It addresses the need to call external drawing APIs directly within DSH conversations and enables agents to generate images.

This plugin is maintained by the user Quqing-future and is licensed under the MIT license.

Core Features

The plugin provides two image generation modes and includes complete configuration and interaction features:

  • Conversational Tools: Supports the generate_image (text-to-image) and image_to_image (image-to-image) tools. They can be invoked directly during conversations, and the generated images are saved to disk and displayed directly in the chat message.
  • Image-to-Image Support: Supports 1 to 10 reference images, with local paths, URLs, or Data URL formats.
  • Settings Panel: Provides a manual image generation panel that can be accessed in Settings under the “Jimeng Image Generation” section.
  • Model Selection: Includes a dropdown list of common Seedream models, supports a “Detect Available Models” button to automatically discover generation models, and also allows custom input.
  • Configurable Options: Allows editing the API Key, baseUrl, text-to-image/image-to-image models, default size, watermark toggle, save directory, and more. Configuration is persisted.
  • Image Saving: Generated images are automatically saved to ~/.dsh/jimeng-images/.
  • Security Mechanism: The API Key is stored on the server side only; the frontend only receives a masked host and does not expose the plaintext key.

Installation and Enabling

Install the plugin from the command line:

dsh plugin --profile web add link:G:\DSHJob\plugins\dsh-jimeng

Or install it manually: add "dsh-jimeng": "file:G:/DSHJob/plugins/dsh-jimeng" to the dependencies in the web profile’s package.json, add "dsh-jimeng" to dsh.profile.bundles, and then run pnpm install.

Usage

After installing and restarting the DSH Web Host, there are two ways to use it:

  1. Conversational Generation: Type an instruction directly in the conversation, for example “Help me draw a whale,” and the AI will automatically call the generate_image tool.
  2. Manual Generation: Go to the settings page, find the “Jimeng Image Generation” section, and manually enter a prompt or select reference images to generate.

Notes

  • Obtaining the API Key: The key must be obtained from the Volcengine Ark console. Be careful to distinguish between “API Key” and account-level AccessKey/SecretKey (AK/SK).
  • Model Versions: In Seedream 3.0, text-to-image and image-to-image use different models; starting with 4.0, the model is unified and can be used for both text-to-image and image-to-image.
  • Image-to-Image Limit: Up to 10 reference images per request (server-side limit). The plugin blocks requests locally if they exceed the limit, so they are not sent.
  • Runtime Environment: Requires Node.js >= 20 and DeepSeek Harness 0.1.x.

This plugin follows DSH’s “everything is a plugin” philosophy and can connect to any service compatible with the OpenAI /images/generations endpoint (such as SiliconFlow or a custom proxy).

The plugin homepage and source code are available at: https://github.com/Quqing-future/dsh-jimeng