One CLI to rule all GPU clouds.
Search pricing across 10 providers, spin up a GPU in seconds, sync your workspace, and track every dollar.

$ swm gpus -g h200 --max-price 4

  Live GPU Availability
┏━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━┳━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━┓
┃ Provider ┃ GPU              ┃ VRAM   ┃ $/hr     ┃ Stock   ┃
┡━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━┩
│ vastai   │ NVIDIA H200      │ 141 GB │ $2.89/hr │ 12 avl  │
│ runpod   │ NVIDIA H200      │ 141 GB │ $3.49/hr │ High    │
│ lambda   │ NVIDIA H200      │ 141 GB │ $3.99/hr │ 4 avl   │
│ vultr    │ NVIDIA H200      │ 141 GB │ $3.88/hr │ 8 avl   │
└──────────┴──────────────────┴────────┴──────────┴─────────┘

Install

# macOS (Homebrew)
brew tap swm-gpu/swm && brew install swm

# Python (3.11+)
pipx install swm-gpu

# From source
git clone https://github.com/swm-gpu/swm.git && cd swm && pip install -e .

Quick Start

# 1. Add your API key
swm config set runpod.api_key <your-key>

# 2. Find a GPU
swm gpus -g h200

# 3. Create a pod
swm pod create -p runpod -g h200 -n my-session

# 4. Install a framework
swm setup install vllm runpod:<id>

# 5. Done — saves workspace and terminates
swm pod down runpod:<id>

Or just ask your agent

Don't want to learn the CLI? Install the SKILL.md and your AI agent manages GPUs for you:

# Universal (works with Cursor, Copilot, Windsurf, Amp, Devin)
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

Works with Cursor, Claude Code, Codex, Copilot, Windsurf, Amp, Devin, and any agent that can run shell commands.

Supported Providers

Provider	GPU Search	Provision	Stop/Resume	Billing API
RunPod	Live	Yes	Yes	Full
Vast.ai	Live	Yes	Yes	Full
Lambda Labs	Live	Yes	—	—
Vultr	Live	Yes	Yes	—
TensorDock	Live	Yes	Yes	—
FluidStack	Live	Yes	Yes	—
AWS (EC2)	Live	Yes	Yes	—
GCP (Compute)	Live	Yes	Yes	—
Azure	Live	Yes	Yes	—
CoreWeave	Live	Yes	Yes	—

Key Features

GPU Search & Provisioning

swm gpus                            # all GPUs, all providers
swm gpus -g h200 -c 4              # 4×H200 configs
swm gpus --max-price 4 --secure    # under $4/hr, certified clouds
swm pod create -p runpod -g h200 -n train
swm pod down runpod:<id>            # sync + terminate

Workspace Sync

Your /workspace directory follows you across clouds via S3-compatible storage (Backblaze B2, Amazon S3, Google GCS).

swm sync push runpod:<id>           # pod → storage
swm sync pull runpod:<id>           # storage → pod
swm sync watch runpod:<id>          # auto-push on file changes

Three-tier smart sync: inotify watcher tracks changes, incremental push uploads only what changed, tar mode packs 600k small files into one S3 object.

Frameworks

swm setup install vllm runpod:<id>       # vLLM inference server
swm setup install open-webui runpod:<id> # Open WebUI chat interface
swm setup install comfyui runpod:<id>    # ComfyUI image generation
swm setup install axolotl runpod:<id>    # Axolotl fine-tuning
swm setup install ollama runpod:<id>     # Ollama model runner
swm setup install swarmui runpod:<id>    # SwarmUI
swm setup install llm-studio runpod:<id> # H2O LLM Studio

Auto-detects GPU count for tensor parallelism, opens SSH tunnels for unexposed ports, probes health endpoints.

Lifecycle Guard

Monitors SSH sessions, GPU utilization, filesystem writes, and active processes. If nothing's happening, it saves your workspace and terminates the pod.

swm guard enable runpod:<id> --policy auto-down --idle 30m
swm guard list

No more $96 overnight H100 bills.

Cost Tracking

swm costs live                      # running cost right now
swm costs summary                   # spending breakdown
swm costs reconcile                 # verify against provider billing APIs
swm costs budget set 100            # $100/month alert

Model Management

swm models search qwen3             # search HuggingFace Hub
swm models pull runpod:<id> Qwen/Qwen3-235B
swm models set runpod:<id> Qwen/Qwen3-235B  # hot-swap vLLM model

How It Works

Everything happens over SSH. No agents on the pod. No custom images. No webhooks.

┌──────────┐       SSH        ┌─────────────┐       S3 API      ┌───────────┐
│ Your Mac │ ───────────────> │  GPU Pod    │ ────────────────> │ B2 / S3   │
│   swm    │  exec, scp      │  (any       │  s5cmd sync       │ / GCS     │
│          │ <─────────────── │   provider) │ <──────────────── │(workspace)│
└──────────┘                  └─────────────┘                   └───────────┘

Credentials are never stored on the pod. Storage keys are passed as transient environment variables per command.

Security

SSH key authentication only — no passwords stored anywhere
No credentials on pods — storage keys passed transiently, never written to disk
Non-destructive syncs — files are never deleted from your storage bucket
Secure cloud default — swm pod create defaults to SOC 2 / HIPAA certified data centers

Documentation

Full docs at swmgpu.com.

Page	Description
Getting Started (CLI)	Install and create your first pod in 5 minutes
Getting Started (Agent)	Let your AI agent manage GPUs for you
Configuration	All config keys for providers and storage
Command Reference	Full reference for every swm command
Core Concepts	Providers, workspaces, frameworks, lifecycle guard

Requirements

macOS or Linux
Python 3.11+ (if not using Homebrew binary)
SSH client (ssh, scp)
An account with at least one GPU provider

License

Licensed under the Apache License, Version 2.0.

NAME

SYNOPSIS

INFO

DESCRIPTION

README