Introduction¶

DeepSeek Harness (DSH) uses a plugin-based architecture to extend its capabilities. This plugin (internal name: Agnes Creative Workstation) is designed to integrate the free image-generation and video-generation large models from Agnes AI, providing DSH with visual creative tools and pipelines.

Core Features¶

The plugin is mainly divided into two categories: model inference and content creation:
* Image and video generation: Supports text-to-image (Image 2.5 Flash, 1K-4K sizes), image-to-image (style transfer, scene transformation, multi-image composition), text-to-video (Video 2.5 Flash, 4-12 seconds), and image-to-video (first-frame control, multimodal references).
* Content production pipelines: Includes a short drama pipeline (script import, storyboard generation, automatic compositing, subtitle burn-in), digital human talking-head video (TTS voiceover, segment-by-segment video generation), MTV production, and automatic scene segmentation and appearance locking for background videos.
* Creative toolbox: Provides novel tools (long-form splitting, removing the AI tone), novel cover generation, Prompt Expert (Chinese prompt expansion), video parsing (scene cutting, VPT generation), and a multi-capability treasure box (video download, TTS).
* Workflow orchestration: Infinite canvas (node-based orchestration, parameter passing through connections), Creative Workbench (Oh Story workflow integration), and WeChat Official Account draft management.

Installation and Configuration¶

Run the installation command inside the DSH container:

npm install /root/软件/dsh-agnes-studio

Edit package.json, and add the plugin references to dsh.profile.bundles and dependencies. If @oh-story/dsh was previously installed, remove it from both locations above to avoid duplicate skill registration. Restart the DSH container after installation.

Environment Dependencies¶

Some features depend on the host machine environment:
* ffmpeg + ffprobe: Used for video concatenation, subtitle burn-in, and audio-video compositing. If missing, the top of the panel will show installation instructions.
* Chinese fonts: Used for subtitle rendering. If not installed, it will return an explicit error instead of silently outputting tofu blocks.
* MIMO_API_KEY: Used for TTS voiceover in digital human talking-head videos. It must be configured in credentials or environment variables.

Usage¶

1. Image/video generation: Open the plugin from the sidebar, enter a prompt (or use the “Try” examples), select the model and size, and click Generate.
2. Short drama pipeline: Import the script → generate the storyboard → click “Generate All Shots with One Click” → compose the final video (optional subtitle burn-in) → download.
3. Digital human talking-head video: Switch to the “🎙 Talking Head” tab, enter the script, select the character image, start generation, then preview or download.
4. Infinite canvas: Switch to “🕸 Canvas”, pan and operate by dragging and zooming with the mouse wheel, use node connections to pass reference images, and initiate generation directly on a node.
5. Novel cover: Switch to “📕 Cover”, upload TXT/DOCX or manually enter the title, select a style, and generate with one click.

Technical Details and Notes¶

• Host/client separation: Instead of messy code-level merging, the agnes host-side component and the oh-story host-side component are combined at lib/host-entry.js; client registration uses a private ID to avoid conflicts. Changes to the host side require restarting DSH, while changes to the client side require refreshing the page.
• State management: The canvas state is stored in the browser’s localStorage and does not depend on the backend.
• API Key security: All API Keys are handled by the host-side proxy; the browser side does not expose plaintext API Keys.
• Panel size: Supports switching among compact, standard, and full-screen sizes; settings are stored locally.

Summary¶

This plugin deeply integrates the free large-model capabilities of Agnes AI with the DSH plugin ecosystem, providing end-to-end tools from generation to composition. It is suitable for users who need film and video production, short drama development, and content-assisted creation. For the source code and documentation, see GitHub.