vastai-skill

Let your AI Agents have access to GPU on Vast.ai.

An agent rents a GPU with the native vastai CLI, works on it over plain ssh/rsync, and destroys it when done. This repo provides the two missing pieces:

skills/vastai-gpu/SKILL.md — the skill that teaches an agent the full workflow: find/rent → connect → sync/run → transfer artifacts → destroy.
vastai-connect <instance_id> — the one gap the native CLI leaves: blocks until a freshly rented instance is actually SSH-reachable, then writes a Host alias to its own ~/.ssh/vastai.conf (auto-included from ~/.ssh/config, which is never touched again). Optionally opens VS Code/Cursor with --ide.

Setup

# 1. Native vastai CLI + API key (from https://cloud.vast.ai/manage-keys/)
uv tool install vastai
vastai set api-key <key>

# 2. Register your SSH public key at https://cloud.vast.ai/manage-keys/

# 3. This tool
uv tool install git+https://github.com/Yusuke710/vastai-skill.git

# 4. The skill — same SKILL.md works for both agents; symlink it where yours looks.
#    Claude Code:
ln -s "$(pwd)/skills/vastai-gpu" ~/.claude/skills/vastai-gpu
#    Codex:
ln -s "$(pwd)/skills/vastai-gpu" ~/.agents/skills/vastai-gpu
# or ask your agent to add this skill

Then ask your agent for something that needs a GPU. e.g. Train a model on GPU etc

The workflow (what the agent does)

1. Find and rent

vastai search offers 'gpu_name=RTX_3060 num_gpus=1 reliability>0.98' -o 'dph' --raw
vastai create instance <offer_id> --image vastai/pytorch:latest --disk 30 --ssh --raw

2. Wait until reachable

vastai-connect <instance_id> --alias vast-gpu        # blocks, then writes SSH alias

3. Work over plain ssh/rsync. This is where you iteratively run experiments on a GPU

rsync -az --filter=':- .gitignore' ./ vast-gpu:/workspace/proj/
ssh vast-gpu 'cd /workspace/proj && uv sync && uv run python train.py'
rsync -az vast-gpu:/workspace/proj/outputs/ ./outputs/

4. Destroy the instance

vastai destroy instance <instance_id>                # destroy, never stop

Details (long-running jobs, GPU verification, failure handling, cost rules) live in the skill.

Connecting your IDE (optional)

Since the instance gets a plain SSH alias, VS Code or Cursor can attach to it with their Remote - SSH extension — pick the alias (e.g. vast-gpu) from the remote host list, or let the tool open it for you:

vastai-connect <instance_id> --ide code     # or --ide cursor

License

MIT

Vastai Skill

README