quackd
A robot control CLI that runs on your computer and uses Ollama, vLLM or cloud models to plan tasks across robot adapters.
quackd lets robot builders give language-based goals to robots with existing movement and manipulation skills. It runs the planning on a computer while the robot's own controllers carry out the actions. Real-hardware use covers the LeRobot SO-101 arm; the other adapters have only run in simulators or mocks.
- Goal-based control: Send a goal from the terminal or a Claude conversation, with task rules that restrict available skills, limit steps and require approval for selected actions.
- Robot adapters: Work with Microduck, Open Duck Mini, LeRobot arms, XLeRobot, AlohaMini, ToddlerBot and ROS 2 bases, using the capabilities each robot actually provides.
- Shared tasks: Give several robots one goal and let their individual agents divide the work using information about each other's capabilities.
- Learned arm policies: An SO-101 arm can delegate picking and manipulation to a policy server, with those actions subject to confirmation.
- Task memory: Keep notes and recent outcomes for each robot so later runs can draw on earlier attempts.
Local LLM choices include Ollama and vLLM. Claude, OpenAI, Gemini and Grok use external model services. Camera-based observation depends on the robot's capabilities and camera connection. The Python project is open source under Apache 2.0.
