Run an agent

The commands on this page call paid model APIs.

Claude Code#

Install Claude Code so that claude is on your PATH. Set an Anthropic API key, or a subscription token from claude setup-token:

export ANTHROPIC_API_KEY=...
# or, with a Claude subscription:
export CLAUDE_CODE_OAUTH_TOKEN=...
robouse run \
  --task metaworld-push \
  --harness claude-code \
  --model claude-sonnet-5 \
  --out runs/try

metaworld-push needs the Meta-World extra: pip install 'robouse[metaworld]>=0.1.1'.

Codex#

Install Codex so that codex is on your PATH. Set OPENAI_API_KEY, or log in once with codex login:

robouse run \
  --task metaworld-push \
  --harness codex \
  --model gpt-6-astra \
  --out runs/try

What you get#

The command prints the same summary as the reference solution in the Quickstart, and writes the trial folder. To see what the agent did:

• agent/trajectory.json: its messages, reasoning and commands, step by step.
• episode/trace.jsonl: every robo request it made and the response.
• artifacts/recording.mp4: the video.

A trial that could not start (a missing credential, a harness not on PATH, an episode server that did not start) has "exception" set in the summary and in result.json; it is not a score. Errors inside the harness, such as a failed API call, end the episode and score 0.

Options#

• --timeout 600: the agent's wall-clock limit in seconds, instead of the task's own.
• --seed 3: another object and goal layout, where the task's backend supports it.
• Other harnesses and what they need: Harnesses.