CLI commands¶

Install the Python CLI package first, then authenticate with Authentication. For BYOC cluster flags, see BYOC CLI.

Authentication¶

Command	Shown on
`arena login`	Profile (alternative to a PAT)

arena login

Run once per machine (or again after session expiry). The arena agent generate snippet assumes you are already authenticated.

Set ARENA_API_KEY instead when you cannot open a browser. Profile does not print a separate arena subcommand for keys; create the PAT in the UI and export the variable.

Inference¶

The Arena CLI tab appears under Deployment code only for LLM deployments, after status is Deployed. It shows a single arena agent generate command:

arena agent generate YOUR_DEPLOYMENT_NAME --prompt "Your prompt here"

YOUR_DEPLOYMENT_NAME: deployment name from the Agents list (same string as in the Arena client tab).
--prompt: the text you send. The command streams the completion to your terminal.

Authenticate first with arena login or ARENA_API_KEY (see Authentication).

Classic RL and supervised deployments¶

Gym, PettingZoo, recurrent PPO, and supervised (predict) deployments do not show an Arena CLI tab. Use the Arena client tab (ArenaClient + open_inference_agent) for a rollout loop, or wire Manual HTTP yourself.

Bring your own compute¶

Bring your own compute prints an arena cluster register … --install --no-write command when you open Install on a cluster row:

arena cluster register --name 'corp-k8s' --install --no-write

arena cluster also covers token rotation, unregistering, and cloud provisioning. Flags and the rest of the subcommands: BYOC CLI.

Projects, datasets, environments, and experiments¶

Action	Command
Create Advanced Training project	`arena projects create llm-work --description "Advanced Training" --llm-based`
Create Classic RL project	`arena projects create classic-demo --description "Classic RL"`
Create SFT dataset	`arena datasets create support-tickets-sft --category sft --column-mapping '{...}' --file ./sft.csv`
Create preference dataset	`arena datasets create assistant-prefs --category preference --column-mapping '{...}' --file ./prefs.csv`
Create reasoning dataset	`arena datasets create math-word-problems --category reasoning --column-mapping '{...}' --file ./reasoning.csv`
Upload and validate a classic gym	`arena env validate my-env --version v1 --source ./my_env.tar.gz`
Upload and validate a language gym	`arena env validate ticket-sim --version v1 --source ./ticket_sim.tar.gz --language-based`
List gym entrypoints	`arena env entrypoints ticket-sim --version v1`
List compute	`arena resources list`
Submit a job	`arena experiments submit ./run-spec.json --project classic-demo --experiment-name ppo-cartpole --resource-id a2-highgpu-1g --num-nodes 2`
Submit a reasoning job	same, plus `--reward-file ./reward.py`
List experiments	`arena experiments list --project llm-work`
List checkpoints	`arena experiments checkpoints sft-run`

--llm-based is Advanced Training. Classic RL is the default. --category is sft, preference, or reasoning. Mapping keys are Arena names; values are column names in your file. --file is a local CSV or Parquet file, or a Hugging Face parquet folder.

Save a default spec from Python as run-spec.json. Copy a resource id from arena resources list, then submit:

import json
from pathlib import Path

from agilerl.arena import ArenaClient

with ArenaClient() as client:
    spec = client.get_default_run_spec(algorithm="PPO", gym_env="CartPole")
    Path("run-spec.json").write_text(json.dumps(spec, indent=2))

arena projects create classic-demo --description "Classic RL"
arena resources list
arena experiments submit ./run-spec.json --project classic-demo --experiment-name ppo-cartpole --resource-id a2-highgpu-1g --num-nodes 2

arena datasets create support-tickets-sft --category sft --column-mapping '{"prompt":"instruction","target":"response"}' --file ./sft.csv
arena env validate my-env --version v1 --source ./my_env.tar.gz
arena experiments list --project llm-work
arena experiments checkpoints sft-run

Non-tabular files still use the web UI.