Auto-discovered marketplace from liorz/vastai-claude-skill
npx claudepluginhub liorz/vastai-claude-skillInterface with Vast.ai GPU marketplace — search offers, launch instances, and manage GPU workloads
LiorZA Claude Code plugin that lets you rent GPUs, launch instances, run jobs, and manage your entire Vast.ai workflow through natural language.
The headline feature is /run-job — an autonomous agent that searches for a GPU, launches it, uploads your files, runs your job, downloads results, and destroys the instance when done. You describe the job; it handles the rest.
Add the marketplace and install the plugin:
/plugin marketplace add LiorZ/vastai-claude-skill
/plugin install vastai@vastai-skills
To install at a specific scope:
/plugin install vastai@vastai-skills --scope user # personal (default)
/plugin install vastai@vastai-skills --scope project # shared with team (committed to repo)
/plugin install vastai@vastai-skills --scope local # project-local, gitignored
For local development:
claude --plugin-dir /path/to/vastai-claude-skill
pip install vastai
vastai set api-key <YOUR_KEY>
vastai create ssh-key "$(cat ~/.ssh/id_ed25519.pub)"
If you're not sure about your setup, run /vastai-setup after installing.
/run-job — Autonomous Job RunnerAn agent that handles the full lifecycle of a GPU job:
/run-job fine-tune a LoRA adapter on my dataset in ./data using an A100, pytorch image
/run-job run python train.py on a 4090, upload the ./src directory, download /root/output when done
/run-job use template abc123 on 8x H100, run bash /root/train.sh, budget $8/hr
/search-gpus — Find GPU OffersNatural language GPU search. Translates your requirements into Vast.ai query syntax.
/search-gpus cheap 4090 in the US
/search-gpus 8x H100 with NVLink for distributed training
/search-gpus any GPU with 80GB+ VRAM under $2/hr
/launch-instance — Create an InstanceGuided instance creation with cost confirmation before billing starts.
/launch-instance offer 12345678 with pytorch image and 100GB disk
/launch-instance RTX 4090 with jupyter lab
/manage-instances — Control Running InstancesShow, start, stop, destroy, SSH, logs, file transfer, snapshots, and scheduled operations.
/manage-instances show all my instances
/manage-instances get SSH command for instance 12345
/manage-instances destroy all stopped instances
/manage-volumes — Persistent StorageSearch, create, delete, clone, and attach volumes that persist across instance lifecycles.
/manage-volumes create a 100GB volume in the US
/manage-volumes list my volumes
/autoscale — Production DeploymentsManage endpoints and auto-scaling worker groups for inference serving.
/autoscale create an endpoint with 4090 workers for my vLLM service
/autoscale show worker group logs
/vastai-setup — First-Time SetupVerify CLI installation, configure API key, register SSH keys, and test connectivity.
The vastai skill (not user-invocable) is automatically loaded when you discuss Vast.ai topics. It provides Claude with comprehensive knowledge of all 124+ CLI commands — instances, volumes, templates, clusters, overlays, teams, billing, hosting, SSH keys, API keys, environment variables, and more.
| Skill | Invocation | Auto-triggers | Runs as agent |
|---|---|---|---|
vastai | Background only | Yes (GPU/vast.ai topics) | No |
search-gpus | /search-gpus | Yes | No |
launch-instance | /launch-instance | No (side effects) | No |
manage-instances | /manage-instances | Yes | No |
manage-volumes | /manage-volumes | Yes | No |
autoscale | /autoscale | Yes | No |
run-job | /run-job | No (side effects) | Yes (forked agent) |
vastai-setup | /vastai-setup | No | No |
Skills with side effects (creating instances, spending money) require explicit /command invocation — they won't trigger automatically.
MIT