Supported

Robo Use 0.4.0.

Models and harnesses

Model · harness	Needs
`baseten/zai-org/GLM-5.3` · `claude-code`, `codex`, `mini-swe-agent`, `dimcode`, `show-harness`	`BASETEN_API_KEY`; Python 3.12 or newer and `pip install 'robouse[llm]'` (the model gateway)
`baseten/moonshotai/Kimi-K3` · `claude-code`	`BASETEN_API_KEY`; Python 3.12 or newer and `pip install 'robouse[llm]'`
`baseten/zai-org/GLM-5.3-Flash` · `show-harness`	`BASETEN_API_KEY`; Python 3.12 or newer and `pip install 'robouse[llm]'`
`claude-sonnet-5` · `claude-code`	`ANTHROPIC_API_KEY` or `CLAUDE_CODE_OAUTH_TOKEN`
`gpt-6-astra` · `codex`	`OPENAI_API_KEY` or a `codex login`
VLA presets · `vla`	a policy server, usually on a GPU
`allenai/MolmoAct2-LIBERO-LeRobot` · `molmoact2`	a policy server on a GPU

Platforms

Platform	Notes
macOS on Apple silicon	Python 3.11 to 3.13; provider-prefixed models need 3.12 or newer
Linux arm64	without a display, OSMesa (`libosmesa6`) or EGL (`libegl1 libgl1`): Robo Use picks whichever renders, and `MUJOCO_GL`, if set, overrides the choice (Quickstart); bubblewrap for model harnesses; the `robosuite` extra needs a C compiler
Python 3.14	not supported: MuJoCo 3.3.0 and Meta-World have no build for it

Isolation

Mode	Isolation
`robouse run` on macOS	model harnesses run under Seatbelt (Run agents and VLAs)
`robouse run` on Linux	model harnesses run under bubblewrap
`bench eval run -d`	agent and simulator in separate Docker containers

Formats

Tasks: BenchFlow task folders (task.md, schema 1.3).
Trials: BenchFlow's trial layout (config.json, result.json, verifier/reward.txt), with the agent's steps in ATIF v1.7 (agent/trajectory.json).