Microduck Agent Plugin

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/

Install

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.

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):

npx skills add acnlabs/microduck-plugin@microduck-skill -g -y

Or point it at skills/microduck-skill after a clone.

ClawHub listing for microduck-skill is still pending review. Use GitHub / npx skills add until it is public.

Worked examples (Hub, not bundled)

Walk and bow are not ONNX files in this repo. They are community Hub graphs this plugin's loop produced. Not official Pollen policies.

Policy	Hub	kind
walk	https://huggingface.co/neil-jo/microduck-walk	perpetual gait
polite-bow	https://huggingface.co/neil-jo/microduck-polite-bow	episodic 4s

Localhost (agent runs control.sh; do not paste these for a human to type):

$SKILL/scripts/control.sh start --repo neil-jo/microduck-walk --detach
$SKILL/scripts/control.sh pull neil-jo/microduck-polite-bow --as polite_bow
$SKILL/scripts/control.sh do polite_bow

Do not start --repo the bow graph. After publish, print robotctl for this run's repo. Run those commands only if the user owns a Microduck and asks to install.

v0 scope: new behavior → smoke → train (default: Hugging Face Jobs, requires MICRODUCK_HF_NAMESPACE) → official ONNX export → Hub publish → optional Hub preview.mp4 (preview.sh, checkpoint replay) → 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). Preview upload is optional and must not fail publish. 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.