CLI & Python SDK
GPUs from your terminal.
Create instances, copy files, run training scripts and stream logs with one CLI and Python SDK. Same workflow across templates, VMs and deployments.
pip install jarvislabs · Python 3.10+
~/project
$ █
Requires Python 3.10 or newer.
bash$ uv tool install jarvislabsEverything the dashboard does is a flag on jl create, and every flag is a keyword argument in Python.
Create, pause, resume and destroy instances on any GPU in the lineup.
Your data survives pause and resume cycles automatically.
Expose ports for Gradio, FastAPI or any web service.
Run a saved script automatically when the instance starts.
Deploy in India or Europe. More regions coming.
Connect over SSH or copy files to and from the instance. CLI only.
jl run uploads your code, builds a virtual environment with uv, runs the script in the background and streams the logs. When it exits, the instance pauses.
Pausing is the default. Pass --keep to leave the instance running, or --destroy to tear it down when the run finishes.
Claude Code, Cursor and Codex can create instances and run experiments through the same CLI. One command installs the skills.
$ jl setup✓ API key savedBalance: $42.10 · Active instances: 0✓ Skills installed for Claude Code# Now the agent can:$ jl create --gpu A100⠋ Provisioning instance...✓ Instance a100-3k9x ready in 38s$ jl run train.py --on a100-3k9x✓ Training started, streaming logs...Epoch 1/10 loss=2.34 lr=5e-4Epoch 2/10 loss=1.87 lr=5e-4
Saves your API key, shows your balance and installs the skill files for the agents you use.
bash$ jl setupYes. The jarvislabs package on PyPI is free. You pay only for the instances you create: hourly GPU time while they run and storage while they are paused.
Python 3.10 and above, for both the CLI and the SDK.
Run jl setup and paste your API key when prompted, or set the JL_API_KEY environment variable. Create a key in the dashboard under Settings, API keys.
Yes. Set JL_API_KEY in your CI environment and pass --yes to skip confirmation prompts. The CLI and the SDK both work non-interactively, so they suit automated training runs, batch inference and test jobs that need a GPU.
NVIDIA H200 SXM, H100 SXM, RTX Pro 6000, A100-80GB, A100, A30 and L4. Run jl gpus for live availability, VRAM and hourly pricing in each region.
One package, an API key, and your first instance is a command away.