In the development scenarios of DeepSeek Harness (DSH), letting an AI Agent directly show local images, videos, or audio to users is a common requirement. Older solutions usually relied on a sidebar or a new tab, and making streaming output appear in real time was difficult. The dsh-media-inline-preview plugin solves this pain point by rendering media directly in the chat message stream using fence syntax.

What Is This

This is a DSH plugin maintained by seeingrain. It provides three fences (mip-img, mip-video, mip-audio) and covers 36 local media formats. The plugin merges and replaces the previously separate image preview and video/audio fence plugins, achieving true inline rendering so media can be viewed or played without navigating to another page.

Core Features

  • 36 Format Support: Covers 10 image formats, 7 native video formats, 9 native audio formats, and transcoding formats.
  • Three Fences: Uses the mip-img, mip-video, and mip-audio fences to distinguish media types.
  • Image Cards and Lightbox: Images are rendered as thumbnail cards, support click-to-fullscreen viewing, zooming, and dragging.
  • Native Players: Video/audio are rendered directly as native controls, support mounting during streaming output, and require no sidebar.
  • Streaming and Large Files: Supports HTTP Range streaming, keeps memory usage constant for large files, with a 2 GiB limit.
  • Transcoding Fallback: Containers that browsers cannot play (such as mkv and avi) automatically trigger ffmpeg transcoding.
  • Performance and Caching: Supports ETag/304 conditional requests, lazy loading, and content-addressed caching of transcoding results for 7 days.
  • Fences as Markdown: Conversation history natively preserves fences, so media is never lost after a page refresh.

Installation and Enabling

Install the plugin and restart the service for it to take effect.

dsh plugin --profile web add github:seeingrain/dsh-media-inline-preview

After installation, the host needs to be restarted, and the client page can be refreshed to load it.

Typical Usage

1. Zero-Configuration Automatic Rendering

After installation, no configuration is required, and media files appearing in a session are automatically rendered. When generating a reply, the Agent first calls the mip_img / mip_media tools to validate the path and returns a fence; the card or player automatically appears in the message.

2. Manually Written Fences

You can also manually write fence syntax:

{"path":"/绝对路径/图片.png","label":"可选标题"}

{"path":"/绝对路径/视频.mp4"}

{"path":"/绝对路径/音乐.mp3"}

3. Notes

  • Namespace Change: Fence tags are namespaced as mip-* (starting from v1.4.0). Old bare tags (such as img, video, and audio) have been deprecated, are no longer recognized, and remain as plain code blocks.
  • Title Display: The label field is only used for manually written fence validation; the UI title always displays only the file name.
  • Multi-Image Grid: Supports the {"images":[{"path":"..."},{"path":"..."}]} format.

Applicable Scenarios and Limitations

  • Security Restriction: The host must be loopback or listed in webRuntime.trustedHosts; cross-site fetching is blocked.
  • Route Limits: The image route limit is 128 MiB, and the media route limit is 2 GiB.
  • Transcoding Cache: Transcoding results are cached by content address for 7 days.

Ecosystem Context

The philosophy of DSH is “everything is a plugin.” This plugin is a community plugin and has no official relationship with DeepSeek / Huanfang.

Plugin catalog: https://www.skillhub.cn/plugins/seeingrain/dsh-media-inline-preview
Source code: https://github.com/seeingrain/dsh-media-inline-preview