When developing DSH agents, integrating image understanding, generation, or speech processing capabilities typically requires calling multiple external services. Under DSH’s plugin-based architecture, extending functionality directly through plugins is a more reasonable approach. The following section introduces the dsh-minimax-multimodal plugin, which integrates MiniMax’s multimodal capabilities into DSH.

What Is It

This is a plugin that provides image understanding, text-to-image generation, speech synthesis, and speech transcription tools for DeepSeek Harness (DSH). It is maintained by lihanhong2002 and released under the MIT License.

Core Features

The plugin provides the following tools:

  • minimax_understand_image: Reads images in the session workspace and answers questions; the default model is MiniMax-M3.
  • minimax_generate_image: Generates an image from text; the default model is image-01.
  • minimax_speak: Synthesizes text into MP3 audio; the default model is speech-2.8-turbo.
  • minimax_transcribe_audio: Transcribes audio into text; the default model is asr-1.0.
  • minimax_list_voices: Lists available system voice options.

After image generation and speech synthesis succeed, the plugin automatically registers the files in DSH’s delivery records, where they are displayed in the file cards below the reply.

Installation and Activation

  1. Add the local plugin directory in the plugin manager of the DSH desktop client.
  2. Add the plugin using the official CLI command. The command is:
dsh plugin --profile desktop add 'link:<插件目录的绝对路径>'
  1. After installation, enable dsh-minimax-multimodal in the plugin manager.

Typical Usage

In a DSH conversation, simply make a natural-language request to invoke the corresponding features:

  • Text-to-Image: Use MiniMax to generate a 16:9 image of a sunset over the beach.
  • Speech Synthesis: Use MiniMax to read this text as MP3: Hello, world.
  • Image Understanding: Use MiniMax to look at images/screenshot.png and explain this error.
  • Transcription: Use MiniMax to transcribe recording.mp3 into text.

Use Cases and Cautions

  • Environment Requirements: This requires a DSH desktop client version with native tools and file delivery interfaces, as well as Node.js 22 or a newer version.
  • Permissions and Limits: The image understanding and transcription tools read existing files in the session workspace and must comply with DSH file system permissions.
    • Image files are limited to 10 MiB.
    • Audio files are limited to 50 MiB.
    • ASR supports audio up to 500 seconds.
    • Speech synthesis supports up to 3000 characters per request.
  • Configure Key: The plugin reads the key through the host credential service, with the configuration item MINIMAX_CN_API_KEY. You can also set an environment variable with the same name in the host environment. Writing the API Key into source code, configuration patches, conversations, or repositories is strictly prohibited. Exposed keys should be revoked in the MiniMax console and regenerated.

Summary

This plugin uses native tools and file delivery interfaces to seamlessly integrate MiniMax’s multimodal capabilities into DSH workflows. For more details or to submit feedback, visit the GitHub repository or the DSH community catalog.