jarvislabs 0.2.0

CLI and SDK for JarvisLabs.ai GPU cloud

jarvislabs

CLI and Python SDK for managing GPU instances on JarvisLabs.ai.

See the docs on Jarvislabs Docs

Installation

As a CLI tool (recommended)

uv tool install jarvislabs

To upgrade:

uv tool upgrade jarvislabs

As a library

pip install jarvislabs

Or with uv:

uv pip install jarvislabs

Requires Python 3.10+.

Authentication

Get your API key at jarvislabs.ai/settings/api-keys.

jl setup

Or set an environment variable:

export JL_API_KEY="your_api_key"

CLI Quick Start

# See available GPUs and pricing
jl gpus

# Create a container instance (pre-configured with PyTorch, Jupyter, IDE)
jl create --gpu A100 --name "my-instance"

# Create a VM instance (bare-metal SSH access)
jl create --gpu A100-80GB --vm --name "my-vm"

# Create an instance and expose a custom HTTP port
jl create --gpu L4 --http-ports 7860

# SSH into it
jl ssh <machine_id>

# Pause when done (stops compute billing, data persists)
jl pause <machine_id>

# Resume later — optionally with different hardware
jl resume <machine_id> --gpu H100

# Destroy when no longer needed
jl destroy <machine_id>

Managed Runs

Run scripts on GPU instances without manual setup. Code is uploaded, a virtual environment is created (with template packages like torch visible by default), and logs are tracked automatically.

For humans, the default mode stays attached to the run, streams logs, and can auto-pause or auto-destroy the instance after the run finishes. For agents, --json is meant for detached workflows and returns immediately, so use --keep and have the agent pause or destroy the instance after the run.

# Run a training script on a fresh GPU (instance auto-pauses when done)
jl run train.py --gpu L4

# Start a long-running web app on a fresh GPU and expose port 8000
jl run app.py --gpu L4 --http-ports 8000 --keep --no-follow

# Pass script arguments
jl run train.py --gpu L4 -- --epochs 50 --lr 0.001

# Sync a project directory and run a script inside it
jl run . --script train.py --gpu A100 --requirements requirements.txt

# Run on an existing instance
jl run train.py --on <machine_id>

# Check on a run
jl run logs <run_id> --follow
jl run status <run_id>
jl run stop <run_id>

More Commands

jl status                   # Account info and balance
jl templates                # Available framework templates
jl list            # List all instances
jl exec <id> -- nvidia-smi   # Run a command remotely
jl upload <id> ./data        # Upload files
jl download <id> /home/results.csv  # Download files
jl ssh-key add ~/.ssh/id_ed25519.pub --name "my-key"
jl scripts add ./setup.sh --name "install-deps"
jl filesystem create --name "datasets" --storage 200
jl get <id>                  # Shows Jupyter + exposed port URLs

Every command supports --help, --json (machine-readable output), and --yes (skip confirmations).

Python SDK

from jarvislabs import Client

with Client() as client:
    # Create a GPU instance (blocks until running)
    inst = client.instances.create(gpu_type="A100", name="my-run")
    print(f"SSH: {inst.ssh_command}")
    print(f"URL: {inst.url}")

    # When done
    client.instances.pause(inst.machine_id)

from jarvislabs import Client

with Client() as client:
    # List and filter instances
    running = [i for i in client.instances.list() if i.status == "Running"]

    # Check GPU availability and pricing
    for gpu in client.account.gpu_availability():
        print(f"{gpu.gpu_type}: {gpu.num_free_devices} free, ${gpu.price_per_hour}/hr")

    # Manage filesystems
    fs_id = client.filesystems.create(fs_name="data", storage=100)

    # Manage startup scripts
    client.scripts.add(script="#!/bin/bash\npip install wandb", name="setup")

Development

uv pip install -e ".[dev]"
uv run ruff format . && uv run ruff check --fix .
uv run pytest

License

MIT