Introduction¶
DeepSeek Harness (DSH) extends functionality through a plugin-based architecture. For developers or users who want more intuitive and immersive visual feedback in the conversational interface, existing pure text or static displays may often be insufficient. The dsh-live2d-avatar plugin was created to address this visual presentation gap by bringing Live2D character models into DSH’s conversation flow.
Plugin Overview¶
dsh-live2d-avatar is maintained by developer clown139880 and released under the MIT License. The plugin provides DSH with switchable Live2D conversation stages, trajectory displays, and desktop pet modes.
Its core capabilities focus on visual presentation and optional voice interaction. It does not replace DSH’s conversation logic, Agents, memory, tools, or history. The plugin includes a publicly usable Live2D Haru sample model and also supports users importing their own Cubism 2/3+ models.
Core Features¶
Views and Stage¶
The plugin provides three main views: Conversation, Trajectory, and Avatar. The Avatar view provides a Live2D stage for displaying the character.
Model Support¶
- Built-in model: After installation, you can see the built-in Live2D Haru sample model.
- Custom model: Supports users importing their own Cubism 2/3+ models. In the Avatar settings, you can configure the model directory, entry file (such as
.model.jsonor.model3.json), and scale and position. Model files can also be placed directly in the plugin’sassets/models/<name>/directory.
Desktop Pet Mode¶
- Web mode: In regular
dsh web, the desktop pet appears as a draggable overlay inside the page. - Desktop mode: In TokensCowork Desktop, the plugin can create a transparent, frameless, always-on-top standalone window. When the main window is minimized or obscured, the character remains visible on the desktop. If native windows are unavailable, it automatically falls back to the in-page desktop pet.
Optional Features¶
- Expression control: Provides optional expression control prompts, allowing the LLM to output action instructions before the dialogue line. Disabled by default on first installation, with per-conversation authorization supported.
- Voice services: ASR (speech-to-text) and TTS (text-to-speech) are disabled by default, so users who have not configured voice services can still use visual features. To enable them, provide HTTP API addresses. Keys are read from the environment variables of the DSH host.
Installation and Enablement¶
Installing the plugin requires DSH command-line tools.
dsh plugin --profile web add dsh-live2d-avatar
dsh web
After installation, open a non-empty conversation and select “Avatar” in the interface to see the built-in Haru model.
Usage Instructions¶
Configure Model¶
To use a custom model, go to the Avatar settings page. Specify the model directory and the entry file relative path, then adjust scale and position. Because model files are served by the DSH service to the browser, make sure to select a directory containing only dedicated model files to avoid accidentally exposing local private files.
Uninstall Plugin¶
To remove the plugin, run the following command:
dsh plugin --profile web remove dsh-live2d-avatar
Notes¶
- Licensing and Dependencies: The built-in Haru model uses the Live2D Free Material License. Its licensing terms are not included in this project’s MIT License. Please read
assets/models/haru/NOTICE.mdfor details. - WebView2 Dependency: The desktop standalone window feature depends on WebView2. The plugin packages the DLL for
Microsoft.Web.WebView2version1.0.3595.46, but the Windows system still requires the WebView2 Evergreen Runtime. - Privacy and Security:
- API keys are not entered in browser settings and are read only from environment variables.
- Voice services are self-deployed or user-trusted services. The plugin does not judge their privacy policies on behalf of the user.
- Feature scope: This plugin only handles visual presentation and optional voice interaction. It does not process core logic such as conversations or memory.
Summary¶
dsh-live2d-avatar is a DSH plugin focused on Live2D visual experience. By providing a conversation stage, trajectory display, and desktop pet modes, it enhances the sense of presence in user interactions with AI. Whether using the built-in Haru model or importing a custom model, the visual experience is consistent in either Web or TokensCowork Desktop environments.