Skip to content
Get packs

Guide · Agent skill

How to use the SkyPilot Multi-Cloud GPU Orchestration agent skill

Updated · written with AI from the project's own documentation and checked against our catalogue data

This guide explains how to install and use the SkyPilot Multi-Cloud Orchestration skill, an agent skill that teaches an AI agent to run machine learning jobs on GPUs from many cloud providers. It is written for readers who have never used an agent skill. By the end you will know how to install the skill, ask for work in plain language, and what you must set up before anything launches.

SkyPilot Multi-Cloud GPU OrchestrationOrchestra-Research/AI-Research-SKILLs · MIT

Quick start

  1. 1Download and unzip the skill. Download the ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder, or upload it under Settings → Capabilities → Skills in the Claude apps.
  2. 2Install SkyPilot. Install Python with pip, then run pip with the skypilot package at version 0.7.0 or later, including extras for your clouds.
  3. 3Check credentials. Configure an account for each cloud you plan to use and run sky check until those clouds show as enabled.
  4. 4Ask for a job in plain language. Describe the workload, GPU type, node count and whether spot instances are acceptable.
  5. 5Preview the launch. Ask the agent to show the optimizer decision with sky launch --dryrun before any cluster starts.
  6. 6Set autostop and monitor. Configure autostop with teardown, then check sky status and sky jobs queue while the job runs.
  7. 7Tear down. Run sky down on finished clusters so billing stops.

What is an agent skill and what does this one make the agent do?

An agent skill is a folder with a SKILL.md file and optional reference documents that an AI agent loads when a task matches the skill's description. This one, named skypilot-multi-cloud-orchestration, covers SkyPilot, an open-source tool that launches ML jobs on GPUs from many cloud providers using one configuration format.

With the skill installed, the agent can write SkyPilot task files, pick the cheapest cloud or region for a GPU type, run jobs on spot (interruptible) instances with automatic recovery, set up multi-node distributed training, and deploy model serving endpoints.

  • Write task YAML files that name GPUs such as A100:8 or H100:8
  • Choose clouds or regions automatically, or list preferred fallbacks
  • Run jobs on spot instances with managed job recovery and checkpointing
  • Configure multi-node training with SkyPilot environment variables
  • Deploy serving endpoints with sky serve

Install the skill from a ZIP file

Download the ZIP for this skill from this page. Then unzip it into one of two places. For a personal install that works in every project, unzip it into ~/.claude/skills/. For a single project, unzip it into that project's .claude/skills/ folder.

If you use the Claude apps instead of Claude Code, you can upload the skill under Settings → Capabilities → Skills.

  • Personal: ~/.claude/skills/skypilot-multi-cloud-orchestration/
  • Project only: .claude/skills/skypilot-multi-cloud-orchestration/ inside your project
  • Claude apps: Settings → Capabilities → Skills, then upload

Set up SkyPilot and your cloud credentials

The skill needs Python with pip and the skypilot package, version 0.7.0 or later. The source installs SkyPilot with extras for the clouds you use. Then run sky check to confirm which clouds are enabled. Each provider needs its own account and credentials, and a cloud that fails the check will not be used.

If you use file mounts, you also need cloud storage buckets, for example S3 or GCS.

  • Cloud accounts and credentials are required for every provider you plan to use (AWS, GCP, Azure, Kubernetes, Lambda, RunPod and others)
  • The skill does not work without credentials for at least one cloud
pip install "skypilot[aws,gcp,azure,kubernetes]" # Verify cloud credentials sky check

Trigger the skill with natural requests

You do not need to name commands. Describe the job and the constraint, and the agent will match the skill's description. Useful triggers include running training or batch jobs across several clouds, using spot instances with recovery, and cutting GPU cost across providers.

Be specific about hardware, the number of nodes, and whether spot is acceptable. Mention a budget or a teardown rule, because the agent can launch expensive multi-GPU clusters.

  • "Write a SkyPilot task to fine-tune on 8 A100s, using spot instances with recovery."
  • "Find the cheapest cloud and region for 4 H100s and show me a dry run first."
  • "Set up a 4-node training job and make sure the cluster stops after 30 idle minutes."
  • "Deploy my model with SkyPilot Serve and autoscale between 1 and 10 replicas."

A first run: the hello world example

The skill's quick start uses a one-GPU test task. Create hello.yaml with a T4 GPU and a short run block. Launch it as a named cluster, connect over SSH, and terminate it when finished.

Terminating matters. A running cluster keeps billing until you stop or delete it.

resources: accelerators: T4:1 run: | nvidia-smi echo "Hello from SkyPilot!" # Launch: sky launch -c hello hello.yaml # SSH to cluster: ssh hello # Terminate: sky down hello

Control cost and safety

Cost is the main risk. Use sky launch with --dryrun to see the optimizer's choice before paying for it. Set autostop so idle clusters are torn down. Spot instances are cheaper but can be preempted, so use managed jobs and checkpoints for long training runs.

The skill runs scripts and uses the network. It launches resources in your accounts with your credentials, so review the task file before approving a launch.

  • Preview the choice with sky launch task.yaml --dryrun
  • Use autostop with down: true, or sky autostop mycluster -i 30 --down
  • Check what is running with sky status and tear down with sky down
  • Use checkpoints on spot jobs and resume from the latest one
  • Keep API keys in envs or secrets rather than in the YAML file
sky launch task.yaml --dryrun sky autostop mycluster -i 30 --down sky status sky down mycluster

Troubleshoot credentials and permissions

Most first-run failures come from credentials or permissions. If sky check shows a cloud as disabled, configure that provider's CLI and check again. Permission errors usually mean the account lacks the compute, storage or IAM roles SkyPilot needs.

The bundled troubleshooting reference lists the verification commands for each provider, including aws sts get-caller-identity, gcloud auth list and az login, and the IAM roles the source names.

  • Cloud disabled in sky check: configure the provider CLI, then run sky check again
  • AWS: the source lists EC2, S3 and IAM permissions
  • GCP: the source lists Compute Admin and Storage Admin roles
  • Azure: the source says Contributor role on the subscription

What people use it for

  • Fine-tuning an LLM on spot A100s with checkpoints written to a bucket
  • Running a multi-node PyTorch training job with torchrun across several GPU nodes
  • Finding the cheapest cloud and region for a given GPU type before a batch run
  • Deploying an inference endpoint with autoscaling through sky serve
  • Falling back from one cloud to another when a GPU type is unavailable

Frequently asked questions

Is the SkyPilot agent skill free?

The skill is published under the MIT licence. SkyPilot is open-source, but the cloud resources it launches are paid services billed by each provider, so you pay for compute, storage and network use in your own accounts.

What do I need to install the SkyPilot skill?

You need Python with pip, the skypilot package version 0.7.0 or later, and an account with credentials for each cloud you want to use. Run sky check to confirm which clouds are enabled. Cloud storage buckets are needed only if you use file mounts.

Where do I put the skill folder?

Unzip it into ~/.claude/skills/ for a personal install or into a project's .claude/skills/ folder for one project. In the Claude apps, upload it under Settings → Capabilities → Skills.

Can I use SkyPilot without cloud accounts?

Not for the workflows this skill covers. It is built around paid cloud resources and will fail without credentials. If you only need a local GPU, this skill is not the right tool.

Why did sky check show my cloud as disabled?

The usual cause is missing or incorrect credentials. Configure that provider's CLI, such as aws configure, gcloud auth application-default login or az login, then run sky check again. The troubleshooting reference in the skill lists these steps.

How does this compare with Modal, RunPod or Ray?

The skill's own guidance says to use Modal for simpler serverless GPU with a Python-native API, RunPod for single-cloud persistent pods, Kubernetes for existing K8s infrastructure, and Ray for pure Ray-based orchestration. Choose SkyPilot when you want one interface across many clouds with cost optimization and spot recovery.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

This skill suits ML engineers and researchers who rent GPUs from several clouds, want to use spot instances with checkpointing, or need multi-node training without hand-built cluster setup. It is a poor fit for anyone without cloud accounts, for unattended use without budget limits, or for production serving that needs strict uptime, since spot instances can be preempted.

Download the skill