Vast.ai Plugin for Claude Code
A 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.
Install
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
Prerequisites
- Vast.ai CLI installed:
pip install vastai
- API key configured (get one at https://cloud.vast.ai/account/):
vastai set api-key <YOUR_KEY>
- SSH key registered with Vast.ai:
vastai create ssh-key "$(cat ~/.ssh/id_ed25519.pub)"
If you're not sure about your setup, run /vastai-setup after installing.
Skills
/run-job — Autonomous Job Runner
An agent that handles the full lifecycle of a GPU job:
- Parses your requirements (GPU, image/template, files, budget)
- Searches for the cheapest suitable offer
- Confirms cost, then launches the instance
- Waits for SSH readiness
- Uploads local files via tar+gzip streaming
- Runs your job (short jobs inline, long jobs via nohup)
- Monitors progress and reports periodically
- Downloads results via tar+gzip streaming
- Destroys the instance (even on failure)
- Reports summary with runtime and estimated cost
/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 Offers
Natural 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 Instance
Guided 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 Instances
Show, 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 Storage
Search, 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 Deployments
Manage 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 Setup
Verify CLI installation, configure API key, register SSH keys, and test connectivity.
Background Reference
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.
How It Works
| 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.
License
MIT
