vast-ai/vastai-sdk
Vast.ai Python SDK — high-level API for GPU instances, volumes, serverless endpoints, and billing.
npx skills add https://github.com/vast-ai/vast-cli --skill vastai-sdkvastai / vastai_sdk)The vastai package provides a Python SDK for managing GPU instances, volumes, serverless endpoints, and billing on Vast.ai. The vastai_sdk package is a backward-compatibility shim that re-exports vastai.
pip install vastai
For serverless and async support:
pip install "vastai[serverless]"
The SDK reads the API key from ~/.vast_api_key by default. You can also pass it explicitly:
from vastai import VastAI
vast = VastAI() # reads ~/.vast_api_key
vast = VastAI(api_key="YOUR_API_KEY") # explicit key
Get your API key from https://console.vast.ai/manage-keys/
The old vastai_sdk import still works:
from vastai_sdk import VastAI # equivalent to: from vastai import VastAI
from vastai import VastAI
vast = VastAI(api_key=None, server_url=None, retry=3, raw=False, quiet=False)
# List all your instances
instances = vast.show_instances()
# Get a single instance
instance = vast.show_instance(id=12345)
# Search GPU offers
offers = vast.search_offers(query='gpu_name=RTX_4090 num_gpus>=4 reliability>0.99')
# Create an instance from an offer
result = vast.create_instance(id=<offer_id>, image="pytorch/pytorch:latest", disk=50)
# Lifecycle
vast.start_instance(id=12345)
vast.stop_instance(id=12345)
vast.reboot_instance(id=12345)
vast.destroy_instance(id=12345)
# Label an instance
vast.label_instance(id=12345, label="my-training-run")
# Get SSH connection string
ssh_url = vast.ssh_url(id=12345) # returns "ssh -p PORT user@host"
scp_url = vast.scp_url(id=12345) # returns scp-compatible URL
Interruptible (spot) instances are priced below on-demand instances, but can be interrupted at any time by another user with a lower bid. Note: vast.search_offers(type='bid', ...) exposes min_bid, but vast.create_instance(...) defaults to on-demand at dph_total unless you pass bid_price=<floor>. Always pass bid_price after a type='bid' search, otherwise the instance will be rented as an on-demand instance/price instead of as an interruptible.
When outbid, the instance moves to stopped (not destroyed) and storage charges continue. Resume by raising the bid via vast.change_bid(id=..., price=...).
# Search GPU offers (use help(vast.search_offers) for full query syntax)
offers = vast.search_offers(query='gpu_name=RTX_3090 num_gpus>=2')
# Search volume offers
volumes = vast.search_volumes(query='...')
# Search network volumes
net_vols = vast.search_network_volumes()
# Search templates
templates = vast.search_templates()
# Search invoices
invoices = vast.search_invoices()
# copy() takes vast URLs: "[C.|V.]id:path", "cloud_service[.id]:path", or "local:path"
vast.copy("local:./data/", "C.12345:/workspace/data/") # Local → instance
vast.copy("C.12345:/workspace/results/", "local:./out/") # Instance → local
vast.copy("12345:/workspace/", "67890:/workspace/") # Instance → instance (legacy format)
vast.copy("s3.101:/data/", "C.12345:/workspace/") # Cloud service → instance
vast.copy("V.1234:/file", "C.5678:/workspace/") # Volume → instance
vast.copy("V.1234:/file", "s3.101:/workspace/") # Volume → cloud service
vast.cancel_copy(dst_id=12345) # Cancel an in-progress copy
# Cloud sync via a saved cloud connection (see the UI settings page for connection IDs)
vast.cloud_copy(src="./data", dst="s3://bucket/path", instance=12345,
connection=<conn_id>, transfer="Instance To Cloud")
vast.cancel_sync(dst_id=12345)
Volume copy is currently only supported for copying to other volumes, instances, or cloud services, not local. Do not use /root or / as a destination directory — it breaks ssh permissions on the instance and future copies fail. See https://vast.ai/docs/gpu-instances/data-movement#constraints.
# List all deployments
deployments = vast.show_deployments()
# Get a deployment
deployment = vast.show_deployment(id=42)
# Delete a deployment
vast.delete_deployment(id=42)
machines = vast.show_machines()
machine = vast.show_machine(id=10)
vast.list_machine(id=10, price_gpu=0.30)
vast.unlist_machine(id=10)
keys = vast.show_ssh_keys()
vast.create_ssh_key(ssh_key="ssh-rsa AAAA...")
vast.delete_ssh_key(id=5)
members = vast.show_members()
vast.invite_member(email="[email protected]", role="developer")
vast.remove_member(id=7)
SyncClient provides typed, synchronous access to GPU offers and instances.
from vastai import SyncClient
client = SyncClient(api_key="YOUR_API_KEY") # or reads ~/.vast_api_key
# Search offers with structured filters
offers = client.search(
num_gpus=2,
gpu_name="RTX_4090",
min_reliability=0.99,
max_dph_total=2.0,
)
# Create an instance
instance = client.create_instance(
offer_id=<id>,
image="pytorch/pytorch:latest",
disk_gb=50,
)
# List your instances
instances = client.show_instances() # returns list[SyncInstance]
# Destroy an instance
client.destroy_instance(instance_or_id=12345)
AsyncClient provides async access to GPU offers and instances. Use as an async context manager.
import asyncio
from vastai import AsyncClient
async def main():
async with AsyncClient(api_key="YOUR_API_KEY") as client:
# Search offers
offers = await client.search(num_gpus=1, gpu_name="A100")
# Create instance
instance = await client.create_instance(offer_id=<id>, image="ubuntu:22.04")
# List instances
instances = await client.show_instances() # returns list[AsyncInstance]
# Destroy instance
await client.destroy_instance(instance_or_id=instance.id)
asyncio.run(main())
For inference endpoints (requires pip install "vastai[serverless]"):
import asyncio
from vastai import Serverless
async def main():
serverless = Serverless() # reads ~/.vast_api_key
# Get an endpoint
endpoint = await serverless.get_endpoint("my-endpoint")
# Make a request
response = await serverless.request("/v1/completions", {
"model": "Qwen/Qwen3-8B",
"prompt": "Who are you?",
"max_tokens": 100,
"temperature": 0.7,
})
text = response["response"]["choices"][0]["text"]
print(text)
asyncio.run(main())
# Find cheapest 4x RTX 4090 and launch a job
from vastai import VastAI
vast = VastAI()
offers = vast.search_offers(query='gpu_name=RTX_4090 num_gpus=4 reliability>0.99')
cheapest = min(offers, key=lambda o: o['dph_total'])
result = vast.create_instance(id=cheapest['id'], image="pytorch/pytorch:latest", disk=100)
print(f"Launched instance: {result['new_contract']}")
# Use help() to explore method signatures
help(vast.search_offers)
help(vast.create_instance)
Take vast-ai/vastai-sdk from the repository into ~/.claude/skills for personal
use, or into .claude/skills inside a project.
The agent identifies a skill by the name field in its header. Two skills with the
same name cannot sit side by side — one of them will be ignored.
The instructions reference pip.
Without those the skill loads but fails at the first command.