Give an agent a face: the DeepSeek Harness plugin
dsh-plugin-mynah lets a DeepSeek Harness agent drive a Mynah digital human: make it speak to the person watching, wave, interrupt, list who is connected, check health, feed the knowledge base. Repo: github.com/honwee/dsh-plugin-mynah.

Tools
| Tool | Does | Needs admin login |
|---|---|---|
mynah_speak | avatar says text; echo verbatim or chat through Mynah's brain; queues behind the current sentence unless interrupt=true | no |
mynah_action | gesture such as wave | no |
mynah_interrupt | stop talking | no |
mynah_sessions | live sessions and their ids | yes |
mynah_channels | published channels, URLs, available gestures | yes |
mynah_status | health of core / engines / ASR / TTS / DB | yes |
mynah_kb_add | add a text document to a knowledge base | yes |
Install
dsh plugin --profile web add dsh-plugin-mynahThen point it at your instance in the profile's cordis.patch.yml (or MYNAH_URL, MYNAH_ADMIN_URL, MYNAH_USER, MYNAH_PASSWORD in the environment):
- id: mynah
name: dsh-plugin-mynah
config:
baseUrl: https://your-mynah:8443
adminUrl: https://your-mynah:9443
username: admin
password: "********"Open a channel page, click 开始对话, and ask the agent: "find the live Mynah session, wave, introduce yourself, then report system health."
Latency, measured
On the public demo (Qwen realtime brain): request to first cloud audio 0.37–0.52 s, to first lip-synced frame 0.73–1.07 s. Most of what you perceive in an agent demo is the agent's own reasoning between tool calls, not the avatar.
Other agent frameworks
The plugin is a thin wrapper over five HTTP routes; porting it to another framework is an afternoon. See Embed, level 3.
