Open Source Robotics

A curated collection of high-quality open source robotics projects, tools, and software to propel the robotics community forward.

quackd

Command-line agent that lets an LLM drive Microduck, LeRobot arms and other robots

License

Apache-2.0

Overview

quackd is a command line that puts a large language model in charge of the robots you own. Each robot joins through an adapter that declares what the body is and which skills it has; the LLM looks at the camera and the robot's state and picks one of those skills at a time, and the robot's own controllers do the walking and grasping. It drives Pollen Robotics' Microduck, Open Duck Mini v2, a LeRobot SO-101 arm, XLeRobot, AlohaMini, ToddlerBot and any ROS 2 base from a laptop, alone or several together, and can be piloted from Claude over MCP. Apache-2.0, by Rok Benko.

quackd keeps the robot's own skills and the LLM's judgement apart. An adapter publishes a manifest of what a body can do, such as sit, stand, kick and quack on a Microduck, or pick and place on an arm, and a verb that is not in the manifest does not exist. On each turn the pilot LLM sees the camera frame, the state and the last result and chooses one verb with its parameters; quackd checks it against an allowlist, a budget and confirmation rules, runs it, observes the result and asks again until the task is done or judged impossible.

The pilot can be one of twelve cloud vendors' models or a local one through Ollama, vLLM, llama.cpp or LM Studio, and Claude can drive a robot or a whole flock through quackd's MCP server. Several robots can share a task, dividing the work between them, and notes from earlier runs are fed into the next one.

Without a robot you can try it in a 2D cartoon simulator, in MuJoCo running Microduck's own trained walking policy, in the SO-101 arm's simulator, or in a browser demo. The README marks for every robot how far it has really got: on version 0.32, one of the nine bodies, an SO-101 arm, has run on real hardware.

Why it matters

Small open robots like Microduck and Open Duck Mini ship with a handful of trained moves but no way to chain them into a task. quackd gives them an off-board brain without retraining anything: the LLM plans, the robot's own controllers move, and the same commands work across very different bodies. It is also unusually honest about what has run on hardware and what has only run in simulation.

Problem it solves

Each open robot comes with its own SDK and its own fixed behaviours, so asking one to find a ball and kick it, or several to work together, means writing glue code for every body. quackd puts one command line and one rule set in front of all of them, so a goal in plain words becomes a sequence of the robot's real skills.

Key takeaways

  • An LLM picks one of the robot's own skills per step; the robot's controllers do the moving
  • Nine robot adapters, including Microduck, Open Duck Mini v2, an SO-101 arm, XLeRobot and any ROS 2 base
  • Works with twelve cloud LLM vendors or local models, and with Claude over MCP
  • Several robots can share one task, and memory carries notes between runs
  • 2D, MuJoCo and browser simulators let you try it without a robot

Who it is for

  • Owners of small open robots such as Microduck, Open Duck Mini or an SO-101 arm who want them to carry out tasks in plain language
  • Developers building LLM agents that act through real robot skills, including over MCP from Claude
  • Researchers trying multi-robot coordination in simulation before hardware

Limitations

  • Only one of the nine bodies, an SO-101 arm, has run on real hardware (15 and 23 September 2026); Microduck itself has run only in simulation
  • The default simulator is a cartoon that tests the agent loop, not physics; only the Microduck and the arm have physics simulators
  • Success on a solo run is the LLM's own claim, checked against ground truth only in simulation
  • Flocks, LAN discovery and the robot-to-robot passport have run only on one machine or in simulation
  • Every step is an LLM call, so runs cost API time and money unless you use a local model
  • A young project: created in August 2026 and changing quickly

This website requires JavaScript to function properly. Please enable JavaScript to explore our collection of open source robotics projects.