Provides both the CLI and SDK for managing Vast.ai GPU cloud resources, plus a serverless client for endpoint inference.
Linux, macOS, or WSL — install the CLI with no Python required:
curl -fsSL https://vast.ai/install.sh | bashUpdate anytime with vastai update.
Windows, or using the Python SDK — install from PyPI instead:
pip install vastai-
Get your API key from https://cloud.vast.ai/manage-keys/
-
Set your API key:
vastai set api-key YOUR_API_KEY- Test a search:
vastai search offers --limit 3You should see a short list of available GPU offers.
The vastai command provides full access to the Vast.ai platform from your terminal:
vastai search offers 'gpu_name=RTX_4090 num_gpus>=4'
vastai create instance 12345 --image pytorch/pytorch --disk 32 --ssh --direct
vastai show instances
vastai stop instance 12345
vastai destroy instance 12345Run vastai --help for a full list of commands. You can also use --help on any subcommand:
vastai search offers --help
vastai create instance --helpfrom vastai import VastAI
vast = VastAI() # uses VAST_API_KEY env var, or pass api_key="..."
vast.search_offers(query='gpu_name=RTX_4090 num_gpus>=4')
vast.show_instances()
vast.start_instance(id=12345)
vast.stop_instance(id=12345)Use help(vast.search_offers) to view documentation for any method.
- Create the client
from vastai import Serverless
serverless = Serverless() # or, Serverless("YOUR_API_KEY")- Get an endpoint
endpoint = await serverless.get_endpoint("my-endpoint")- Make a request
request_body = {
"model": "Qwen/Qwen3-8B",
"prompt" : "Who are you?",
"max_tokens" : 100,
"temperature" : 0.7
}
response = await serverless.request("/v1/completions", request_body)- Read the response
text = response["response"]["choices"][0]["text"]
print(text)Find more examples in the examples/ directory.
Vast.ai has a skill for AI coding agents (Claude Code, Cursor, Windsurf, Codex, etc.):
npx skills add vast-ai/vast-cliThis installs the Vast.ai skill so your agent can search offers, create instances, and manage GPU workflows directly. See CLI SKILL.md or SDK SKILL.md for the full reference.
This repository is open source. If you find a bug, please open an issue. PRs are welcome.