Getting started

When to use: You want an AI agent to do the work with you, whether you are new to Lightning AI or already use it. Give your coding agent, such as Claude Code, Codex, or Cursor, the Lightning AI skills and ask it to start Studios, run jobs, or deploy models for you. Take one of the other paths when you would rather do it yourself.

1. 1

   Install the Lightning AI skills

   The skills need Python 3.11 or later, with uv or pip. Skills are instruction files your coding agent reads to learn how to use Lightning AI, such as how to start a Studio, run a job, or deploy a model. Pick the tab for the agent you use.

   Claude CodeOther agents

   Run these in your terminal. They work before Claude Code is open, so they also fit scripts and automated setups. The plugin adds extras such as live job progress bars.

   claude plugin marketplace add Lightning-AI/skills
   claude plugin install lightning@lightning-ai

   Already in a Claude Code session? Run /plugin marketplace add Lightning-AI/skills, then /plugin install lightning@lightning-ai.

   The skills.sh CLI installs the same skills into Codex, Cursor, and many other agents. Swap codex for your agent's name, such as cursor, gemini-cli, or github-copilot. It needs Node.js, which you can download from nodejs.org if you don't have it.

   npx skills add Lightning-AI/skills --agent codex --skill '*' --global --yes

   Avoid --all. It installs the skills into every agent on your machine, not only yours.

2. 2

   Install the Lightning CLI

   Your agent uses the lightning CLI to work with Lightning AI. Install it with whichever you have.

   uvpip

   uv tool install lightning-sdk

   pip install lightning-sdk -U

   Inside a Studio, it's already installed.

3. 3

   Sign in

   Sign in once, where your agent runs.

   LocalCIStudio

   On your own machine, lightning login signs you in through your browser and saves it for later commands.

   lightning login

   For CI, cloud agents, or anywhere without a browser. Click your profile picture, open Global settings > Keys, and copy your user ID and API key. Then set them as LIGHTNING_API_KEY and LIGHTNING_USER_ID in that environment's secrets settings.

   Inside a Studio, there's nothing to do. You're already signed in.

4. 4

   Start your agent and send a first prompt

   In your project folder, start your agent, such as claude for Claude Code or codex for Codex. Then send it this. You don't need to mention Lightning AI: the skills bring it in on their own.

   Set this project up on a cloud dev machine: start on a free CPU, copy the code over,
   install its dependencies, and run a quick smoke test.
   Give me the link, and ask before moving to a paid GPU.

   To edit the code yourself, open the Studio link your agent gives you: VS Code runs in the browser. Or connect your local IDE to the Studio to edit its files from your laptop's editor.

When to use: You want something that already runs. A template is a Studio that arrives with its code, data, and environment set up. You clone it into your account, then read it and change it.

1. Pick a template

   Open a template and click Clone to get your own copy, running in your teamspace.

   Browse all templatesTraining, fine-tuning, inference, and more, each one ready to run.

When to use: You would rather do each step yourself than have an agent do it. You start a Studio, bring your code and data, and get it running on a GPU. A Studio works like your laptop, except you can switch the machine underneath from CPU to GPU without reinstalling anything.

1. 1

   Start a Studio

   Your Studio boots on a free CPU, with VSCode and JupyterLab already running in the browser.

   A new Studio lands in your active teamspace, the one selected in the sidebar. If you belong to an organization, check which teamspace that is before you start, since it decides who can see the Studio and who is billed for it.

   Start a Studio

   --create makes the Studio if it does not exist yet, and it starts on a CPU unless you pass --machine. Leave out --teamspace to use your default one.

   On your own machine, install the SDK and sign in first. Inside a Studio, both are already done.

   pip install lightning-sdk -U
   lightning login

   lightning studio start --name my-project --teamspace <OWNER>/<TEAMSPACE> --create

   You only need a GPU to run the training, not to set it up, so install code, data, and dependencies on the free CPU. A training template arrives with the model, data, and code already wired together.

2. 2

   Connect your local IDE

   The browser IDE is ready with no setup. You can also manage the same Studio from your laptop, in which case the code and files still live and run on the Studio.

   Edit in the editor on your laptop while the code, files, and GPU stay on the Studio. Any editor that connects over SSH works.

   1. In the Studio, click SSH in the upper-right corner, then Connect via SSH. Pick your OS and run the setup command it shows on your laptop.
   2. Copy the ssh <address>@ssh.lightning.ai command it shows next.
   3. Paste it into your editor's SSH connection. In VSCode or Cursor that is Connect to Host. Or run it in a terminal.

   If you have the lightning CLI, run lightning ssh configure --name my-project instead and connect to my-project. See Connect your local IDE for more details.

   Edit in the editor on your laptop while the code, files, and GPU stay on the Studio. Any editor that connects over SSH works.

   1. In the Studio, click SSH in the upper-right corner, then Connect via SSH. Pick your OS and run the setup command it shows on your laptop.
   2. Copy the ssh <address>@ssh.lightning.ai command it shows next.
   3. Paste it into your editor's SSH connection. In VSCode or Cursor that is Connect to Host. Or run it in a terminal.

   If you have the lightning CLI, run lightning ssh configure --name my-project instead and connect to my-project. See Connect your local IDE for more details.

3. 3

   Bring your code

   Clone your repo into the Studio and install its dependencies. What you install survives restarts.

   1. Open a terminal in the browser IDE, or in the editor you connected.
   2. Clone your repo, then install its dependencies the way you do on your laptop.

   SSH into the Studio, then clone your repo and install its dependencies.

   lightning studio ssh --name my-project
   
   # then, inside the Studio
   git clone <repository-url>
   cd <repository-directory>
   pip install -r requirements.txt

   Or open the AI copilot in the Studio and ask it to clone the repo and install what it needs.

4. 4

   Add data

   Each teamspace has one Drive, shared by every Studio, job, and deployment in it. Upload your data once and every machine can read it. Only members of the teamspace can see it.

   • The first 10 GB are free.
   • Files your jobs write, such as checkpoints and logs, land on the same Drive.

   1. Open the Drive tab in your teamspace, then drag files or folders in.
   2. Or connect storage you already have, such as S3, GCS, or Cloudflare R2. See Data connections.

   lightning cp copies to your teamspace Drive, and -r copies a whole folder. What you copy shows up under /teamspace/uploads in every Studio in the teamspace.

   lightning cp -r ./data/ lit://<OWNER>/<TEAMSPACE>/uploads/data/

   Inside a Studio, the Drive is at /teamspace/lightning_storage and uploads are in /teamspace/uploads. Write inside a subfolder, since the Drive root is read-only.

5. 5

   Iterate on a GPU

   Switch to a GPU of your choice. Your files and packages come with you. Get one training step running by hand before you start a longer run.

   The Studio stays interactive, so you see and fix problems as they happen.

   1. Open the Studio environment panel in the top bar, then click Switch to GPU and wait until it is ready.
   2. Watch GPU and memory use in the Studio while the step runs.
   3. Fix crashes, batch size, and config in place.

   studio switch moves the Studio to the machine you name. Then SSH in and run one training step by hand.

   lightning studio switch --name my-project --machine H100
   lightning studio ssh --name my-project
   
   # then, inside the Studio
   python train.py --max-steps 1

   Once one step works, submit the same code as a batch job. It runs from a snapshot of your Studio on its own machine, so you have no Dockerfile to write and your Studio stays free.