The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.
MuJoCo, as a Harness agent
Harness agent package for
MuJoCo: describe a robot, a scene, a controller, a policy in the chat pane;
the MuJoCo Viewer pane runs it — MuJoCo's own WebAssembly build, live in the browser, like
simulate: push the robot, drive its actuators, inspect contacts and sensors, replay the recording
frame by frame. Menagerie robots (Unitree Go2, G1, H1, Berkeley Humanoid, Booster T1) come with it;
MJX and MuJoCo Playground are one script away for training. Runs on Claude Code.
Try another future. Open What if…, pin any moment and try lunar gravity, a slippery surface, or a sideways shove. The pane runs both futures from the same state and overlays the original as a wireframe. Scrub time, inspect the paths and body-height chart, and compare measured separation. The controls work on the agent's model and on MJCF you bring yourself. On a narrow pane, the lab docks below the scene so you can still see both outcomes.
Save the experiment to keep the model and assets, full starting state, control tape, changed
conditions and both measured trajectories. The agent can reproduce that file with native MuJoCo
using toolchain/experiments.py, even after the workspace changes, then use your finding to improve
the next controller. CSV exports give you the followed body's positions, separation and contact
counts. These comparisons use recorded controls or held actuator values; they do not rerun a
Python feedback controller or train a policy in the browser.
harness.json— engine, template, skill, toolchain,viewer.use: autonomous/mujoco-viewer.toolchain/setup.sh— one venv on Python 3.12 (uv brings it when the machine has none) with the pinned MuJoCo (VERSIONS), which has no Intel Mac build — Apple Silicon Macs and Linux — and a sparse checkout of the Menagerie robots at a pinned commit;install-training.shadds JAX, MJX and Playground;harness_mujoco.pyloads (servos=turns torque motors into position servos), records and holds poses;verdict.pyjudges the rollout and names the trajectory as the artifact — never the video.- What a rollout is (
record, all underout/):rollout.qpos.json— per frametime,qpos,qvel,ctrl, written while it runs (status: recording → done) — which the pane re-simulates live;rollout.model.xml, the compiled model when it was built withMjSpec, plusmodel_patchfor runtime edits, so the pane runs the model that was simulated;rollout.json, the report; androllout.mp4(optional,video=False) for sharing. skills/mujoco/— the MuJoCo skill (ours).template/— a Go2 standing on position servos, and a pendulum MJCF.
On a Mac, training runs JAX on the CPU — enough for a smoke test, hours for a policy. A GPU machine in Harness's Machines menu is where a real run belongs; the same workspace works there.
Credit and stewardship
MuJoCo is Google DeepMind's — google-deepmind/mujoco,
Apache-2.0 (LICENSE-mujoco) — and so is the MuJoCo Menagerie
(Apache-2.0, with each robot's own licence in its folder — the Unitree models are BSD-3-Clause, Unitree
Robotics). Nothing of either is changed here; MuJoCo is installed from PyPI as released and the
Menagerie is fetched at a pinned commit. This folder is the Harness wrapper — the manifest, a
skill, the helper, the template, the verdict — written by Autonomous to bring MuJoCo into Harness,
on the project's behalf, to bootstrap the catalogue.
If you maintain MuJoCo or the Menagerie and want to own this package, it is yours: open an issue on
OpenHarness and we move this folder into a
repository of yours and point the registry entry at it. Until then: bugs in MuJoCo belong upstream, bugs in the
wrapper belong here, and a newer MuJoCo or Menagerie is a bump of VERSIONS.
harness dsh check . # conformance
harness dsh install "$PWD" --link # this checkout as the installed agent
python3 -m unittest toolchain/test_verdict.py # the verdict, without mujoco
.venv/bin/python -m unittest toolchain/test_record.py # what record() writes for the pane