Introduction¶

The design philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” When developers build agents in DSH, relying solely on text output often lacks an intuitive interactive experience. dsh-plugin-mynah addresses this pain point. It acts as a bridge between DSH and the Mynah real-time digital human stack, allowing agents to control a digital human for real-time speech, interruption, and gesture interaction.

Plugin Positioning¶

This plugin is maintained by honwee and released under the MIT license. It allows DSH agents to directly drive a Mynah real-time digital human (WebRTC-based lip-sync video), enabling text-to-speech, session management, action control, and knowledge base updates.

Core Features¶

The plugin provides a set of tools, divided into real-time interaction tools and management tools:

1. Real-time interaction

   • mynah_speak: Makes the digital human speak in the current session. Supports type=echo (read text verbatim) or type=chat (uses Mynah’s own brain). By default, new sentences are queued unless interrupt=true is set.
   • mynah_interrupt: Forces interruption of the digital human currently speaking.
   • mynah_action: Plays a preset action, such as waving.

2. Session and management

   • mynah_sessions: Lists currently active session IDs for subsequent calls.
   • mynah_channels: Lists published channels and visitor access links.
   • mynah_status: Checks the health status of core services, avatar engine, ASR, TTS, and database.

3. Knowledge base

   • mynah_kb_add: Adds text or Markdown documents to Mynah’s knowledge base.

Installation and Configuration¶

Install the plugin using the official CLI:

dsh plugin --profile web add dsh-plugin-mynah

After installation, the address of the Mynah instance must be specified in the DSH configuration. You can add a configuration block in cordis.patch.yml:

- id: mynah
  name: dsh-plugin-mynah
  config:
    baseUrl: https://your-mynah-host:8443      # 访客/实时端点
    adminUrl: https://your-mynah-host:9443     # 管理后台端点
    username: admin
    password: ********

Alternatively, use environment variables (MYNAH_URL, MYNAH_ADMIN_URL, MYNAH_USER, MYNAH_PASSWORD, etc.) for configuration.

Usage Examples¶

Talk in the Web UI, or run a script in Headless mode:

Scenario 1: Web UI
Enter the prompt in the DSH chat window:

“Find the live Mynah session and have the digital human greet me by name.”

Scenario 2: Headless script

dsh --profile headless "Use mynah_sessions, then make the digital human say: 今天的部署已经完成。"

Runtime Requirements and Notes¶

• Environment dependencies: DeepSeek Harness version must be ≥ 0.1.2-alpha.5, and Node.js version must be ≥ 22.19.
• Mynah instance: Accessible Mynah instance must be deployed. Mynah is a self-hosted digital human stack that uses free EdgeTTS by default, requiring no API key to speak.
• Hardware resources: Mynah runs on WebRTC and the runtime environment is recommended to have approximately 6GB of VRAM (or use remote inference).
• Permission notes: Real-time tools (such as mynah_speak) only require a session ID and no login; management tools (such as mynah_sessions) require administrator login and a cached JWT token.

Conclusion¶

The plugin connects LLM reasoning with digital human interaction through the mynah_* toolchain, giving agents more concrete interaction capabilities. Complete documentation and source code are available in the GitHub repository.