Agent skills that teach AI coding agents (Claude Code, Cursor, and any agent that supports the SKILL.md format) how to use the Lightning AI platform: GPU Studios, batch jobs, model deployments, code-execution sandboxes, the LLM gateway, durable shareable artifact links, and up-front cost estimates for any of it.
| Skill | What it covers |
|---|---|
lightning-studios |
Create, start, stop and manage cloud GPU Studios; switch machines, run commands, transfer files, SSH |
lightning-jobs |
Launch and monitor batch jobs (single and multi-machine) on CPUs/GPUs, stream logs, show a live progress bar with ETA and setback tracking, SSH into running jobs or multi-machine workers, collect artifacts |
lightning-deployments |
Deploy containers/APIs with autoscaling, manage releases, endpoints and auth |
lightning-sandboxes |
Fast ephemeral VMs for safe code execution: run commands, background processes, file I/O, Docker (docker / docker compose) and public port URLs |
lightning-llm-gateway |
Call hosted LLMs (OpenAI, Anthropic, open models) through Lightning's models API |
lightning-artifacts |
Publish a file and get a durable, public lightning.ai/artifacts/<id> link that never expires and renders inline; list, revoke, and delete shares — entirely via the CLI with regular auth |
lightning-cost-estimation |
Quote what a training run, fine-tune or deployment costs: live per-hour GPU/CPU prices for every cloud, spot rates, multi-node fan-out, and Drive storage |
Tutorial: teach a 4B model to call tools on a cloud GPU. It starts from a Mac or a Claude cloud session and one chat message.
All skills are built around the lightning-sdk Python package and its lightning CLI, plus the raw lightning api escape hatch for anything the SDK doesn't wrap.
This repo is also a plugin marketplace, so Claude Code can install all seven skills as one plugin and keep them updated:
/plugin marketplace add Lightning-AI/skills
/plugin install lightning@lightning-aiPlugin skills are namespaced, so they appear as /lightning:lightning-jobs, /lightning:lightning-studios, and so on. They still fire on their own when a task calls for them.
The skills.sh CLI installs into Claude Code, Cursor, Codex, and many other agents:
# interactive: pick skills and target agents
npx skills add Lightning-AI/skills
# install everything without prompts
npx skills add Lightning-AI/skills --all -y
# install a specific skill, e.g. just sandboxes
npx skills add Lightning-AI/skills -s lightning-sandboxes
# user-level (global) instead of the current project
npx skills add Lightning-AI/skills -gOr copy the skill folders straight into your agent's skills directory:
SKILLS="lightning-studios lightning-jobs lightning-deployments lightning-sandboxes lightning-llm-gateway lightning-artifacts lightning-cost-estimation"
# Claude Code (project-level)
mkdir -p .claude/skills && cp -r $SKILLS .claude/skills/
# Claude Code (user-level)
cp -r $SKILLS ~/.claude/skills/- Python with
uvorpip. Skills use thelightningCLI in your current environment, installing or upgradinglightning-sdkthere if it's missing or too old (uv tool install lightning-sdkoruvxare fallbacks) - A Lightning AI account: authenticate with
lightning loginor setLIGHTNING_API_KEY(plusLIGHTNING_USER_ID, optional)
Important
These skills create real, billed resources on your Lightning AI account (Studios, jobs, deployments, sandboxes). They ask before anything that spends money or makes something public, but you are responsible for what runs under your account.
- Skills never guess the organization or teamspace: when more than one is available and none is configured, they ask the user which one to use.
- Skills prefer the documented SDK/CLI surface and fall back to
lightning api <endpoint>for raw REST calls.
Adding or editing a skill? See CONTRIBUTING.md for the SKILL.md
format, the house-style section layout, the conventions above, and how to
live-test every command against a control plane before opening a PR.
- Bugs and feature requests: open an issue
- Security problems: see SECURITY.md, and don't file them as public issues
- Changes between versions: CHANGELOG.md
Apache-2.0. By contributing you agree your contributions are licensed the same way, and to follow the code of conduct.