Skill Overview

The swm SKILL.md is a structured document that teaches AI agents how to provision and manage GPU pods end-to-end. It follows the open Agent Skills standard supported by Cursor, Claude Code, Codex, Copilot, Windsurf, and others.

Install the skill

The fastest way to get started:

mkdir -p .agents/skills/swm-gpu-workflow
curl -sL https://raw.githubusercontent.com/swm-gpu/swm/main/.agents/skills/swm-gpu-workflow/SKILL.md \
  -o .agents/skills/swm-gpu-workflow/SKILL.md

See Supported Platforms for platform-specific install instructions.

The 6 Phases

The skill guides agents through a repeatable workflow:

Phase	What happens
0. State check	Check for existing pods with swm pod list
1. Clarify	Ask user about VRAM needs, provider preference, lifecycle policy
2. Pick GPU	Search live pricing with swm gpus, select cheapest match
3. Provision	Create pod with swm pod create -y, wait for SSH
4. Install	Use swm setup install for built-in frameworks, or swm run for custom tools
5. Verify	Health-check HTTP endpoints, validate GPU memory, check disk space
6. Hand off	Report URL, lifecycle policy, and debugging commands to user

What the agent decides

• GPU class based on VRAM requirements
• Cheapest in-stock provider across all 10 clouds
• Tensor parallelism configuration (auto-detected from GPU count)
• Whether to use built-in swm setup or custom swm run chains
• Lifecycle guard policy and idle timeout

What the agent asks

• VRAM floor needed
• Provider preference (or auto-select cheapest)
• Lifecycle policy (auto-down, auto-stop, remind, manual)
• Idle timeout duration
• Whether to persist workspace to S3

Built-in framework support

The skill covers all 7 built-in frameworks:

Framework	Command	Use case
vLLM	swm setup install vllm	LLM inference server (OpenAI-compatible API)
Open WebUI	swm setup install open-webui	Chat interface for LLMs
Ollama	swm setup install ollama	Local model runner
ComfyUI	swm setup install comfyui	Node-based image generation
SwarmUI	swm setup install swarmui	Image generation UI
Axolotl	swm setup install axolotl	LLM fine-tuning
H2O LLM Studio	swm setup install llm-studio	No-code fine-tuning

For unsupported tools, the skill guides agents through swm run command chains.

Additional capabilities

Beyond the core workflow, the skill covers:

• Lifecycle Guard — swm guard enable to auto-terminate idle pods
• Cost Tracking — swm costs live for real-time spend, swm costs budget set for alerts
• Model Management — swm models search / pull / set for HuggingFace models
• Workspace Sync — swm sync watch for continuous auto-push on file changes

Anti-patterns the skill prevents

• Installing to container disk instead of /workspace (data loss on stop)
• Using pip install -e . for projects with [tool.uv.sources]
• Forgetting -y on swm pod create (hangs in agent context)
• Mixing venvs across frameworks with conflicting torch versions
• Handing off a pod before all health checks pass
• Skipping the state check (creating duplicate pods)