ArenaOS CLI

Discover registered components

These commands print the environments, agents, and plugins registered by the CLI process. IDs shown here are the IDs accepted by run commands.

pnpm arena environments
pnpm arena agents
pnpm arena plugins

Launch an experiment

Choose a world and agent. Multi-participant environments can also receive an opponent. The CLI selects the correct default evaluator and stores the resulting run under .arena/runs.

pnpm arena run royal-chess-v1 --agent royal-greedy --opponent royal-positional
pnpm arena run biocraft-v1 --agent biocraft-researcher
pnpm arena run agent-rumble-v1 --agent rumble-tactician

Control budgets and output

Use step, token, cost, and evaluator options to make runs bounded and repeatable. JSON and quiet modes make the command suitable for scripts and CI.

--max-steps <number>
--max-tokens <number>
--max-cost-usd <number>
--evaluators <id,id>
--json
--quiet

Inspect persisted evidence

List run history, inspect a single record, or render the textual reference replay. Rich environment-specific replays remain available in the web interface.

pnpm arena runs
pnpm arena inspect <run-id>
pnpm arena replay <run-id>

Model-backed runs

When OPENROUTER_API_KEY is configured, model agents appear in the same agent registry as defaults. Pass the selected registered agent ID to --agent or --opponent; credentials never appear in CLI output or run configuration.