Quackd

This open source command line tool lets an LLM control robots. Each robot joins through an adapter that declares what it is and which skills it has, and the model chooses one of those skills at a time, such as look, walk or kick, until the stated goal is met or judged impossible. Balance, gait and grasping stay with the robot's own controllers, while the planning runs on a laptop or in the cloud. It is built for hobbyists and researchers who own or simulate robots. Goals can be given from a terminal, from a task file with the .duck extension, or through a chat with Claude over MCP. A group of robots can be handed one goal, with each robot given its own model and the group dividing the work by announcing what each will do. Safety controls include allowlists, budgets, confirm gates and a kill switch enforced outside the prompt. The site is direct about maturity. It supports seven robot bodies and eleven model providers, but only one real robot, an SO 101 arm, has run on hardware, and the other six adapters are simulated or mocked. Several goals pass ten of ten seeds in the bundled MuJoCo simulator, and a browser demo needs no API key. The software is licensed under Apache 2.0, requires Python 3.11 or later, installs at about 250 MB with no GPU, and is published on PyPI. No price is charged for it. The site also promotes a separate community token, QUACKD, which it states is not the software and has no utility.