https://github.com/acnlabs/microduck-plugin · Agent Plugins 1.0 package for the Microduck train → export → Hub → robot loop.
This is not the robot firmware. That repo is pollen-robotics/microduck. Compatible clients discover this folder via plugin.json, then load the skill microduck-skill.
microduck-plugin/
├── plugin.json
├── LICENSE
├── README.md
└── skills/microduck-skill/
Clone this repository. Add the directory that contains plugin.json to an Agent Plugins client (Cursor, GitHub Copilot, ChatGPT/Codex, VS Code, Kiro). Do not add only skills/microduck-skill if the client supports plugins.
How the client is told to load a plugin directory is client-specific. Two portable options:
git clone https://github.com/acnlabs/microduck-plugin
# Point the client at that folder (the one with plugin.json).If the client only scans Agent Skills (not plugins), point it at skills/microduck-skill.
v0 scope: new behavior → smoke → train (default: Hugging Face Jobs, requires MICRODUCK_HF_NAMESPACE) → official ONNX export → Hub publish → localhost sim control → deploy. Optional wandb login turns on webpage curves. Jobs cannot record --video (no OpenGL). play.sh is local Viser + local mp4 (training checkpoint). The agent runs control.sh (search / start / twist / do) itself — do not expect the human to type those. After publish the skill always prints robotctl for that repo. It runs those commands only if you explicitly ask to install on a Microduck you own.
Package CI (no GPU): python3 scripts/ci_check.py
Ready-check (needs uv and a Hugging Face login):
./skills/microduck-skill/scripts/doctor.sh --clone --sync export MICRODUCK_RL_ROOT=... export MICRODUCK_HF_NAMESPACE=<hf-user-or-org>
Upstream:
- Runtime: https://github.com/pollen-robotics/microduck
- Training: https://github.com/pollen-robotics/microduck_rl (
AGENTS.md)
Related (local GPU box / CLI, not this Jobs-first plugin): microduck-mcp, microduck-cli.