Account access.
Create keys in the console. CLI access is enabled per account — contact us about access for your workflow. This page describes installation, authentication and the available commands.
Install#
$ curl -fsSL https://cli.onlinegpu.io/install.sh | sh $ spot version spot 0.4.0
A single static binary, no runtime to install. Windows users can run it under WSL. The installer never needs root: it drops the binary in ~/.local/bin.
Authenticate#
Create a key in the console under API keys, then:
$ spot auth login --key sg_live_xxxxxxxxxxxxxxxx Signed in as you@example.com · balance $128.40 # Or keep it out of the config file entirely: $ export SPOT_API_KEY=sg_live_xxxxxxxxxxxxxxxx
The key is stored in ~/.config/onlinegpu/config.toml with mode 0600. SPOT_API_KEY takes precedence, which is what CI should use.
spot launch#
$ spot launch --gpu h100-sxm --count 1 --tier spot \ --region us-east --template pytorch --volume train-data --relaunch Instance i-7f3a2c H100 ×1 us-east spot @ $0.55/h ssh root@i-7f3a2c.onlinegpu.io ✓ reachable in 41s
| Flag | Default | What it does |
|---|---|---|
--gpu <slug> | required | A model from spot gpus, for example h100-sxm |
--count <n> | 1 | GPUs in the node; must be a size the model offers |
--tier spot|ondemand | spot | Whether the instance may be reclaimed |
--region <code> | first with capacity | us-east, us-west, eu-central |
--template <slug> | pytorch | A template, or --image ghcr.io/you/img:tag for your own |
--volume <name|ref> | none | Attach a persistent volume at /mnt/vol |
--disk <gb> | none | Create a volume of that size and attach it |
--relaunch | off | Queue for the next free capacity after a reclaim |
--name <text> | none | A label for your own reference, 60 characters |
--json | off | Print the instance object instead of a table |
--wait | on | Block until the instance is reachable; --no-wait returns immediately |
spot ls, spot show#
$ spot ls REF GPU TIER REGION STATE UPTIME COST i-7f3a2c H100 ×1 spot us-east running 02:14 $1.23i-9b10e4 RTX 4090 spot eu-central stopped — $0.13 $ spot show i-7f3a2c --json | jq .volume { "ref": "vol-4b19c7", "mount": "/mnt/vol", "size_gb": 500 }
spot watch#
Follows an instance and runs a command when a reclaim notice arrives — the CLI equivalent of polling the metadata endpoint from inside the machine.
$ spot watch i-7f3a2c --on-notice ./drain.sh --on-stop ./notify.sh 18:04:12 running $0.23 so far 18:12:40 notice reclaim in 02:00 → ./drain.sh 18:14:40 stopped volume vol-4b19c7 kept, address kept 18:18:50 relaunched i-9b10e4 spot @ $0.55/h
spot stop, spot rm#
$ spot stop i-7f3a2c # stops billing, keeps disk, IP and configuration $ spot start i-7f3a2c # same instance, next free capacity of that model $ spot rm i-7f3a2c # terminate; the volume survives until you delete it
spot volume#
$ spot volume create train-data --size 500 --region us-east vol-4b19c7 500 GB us-east $40.00/month $ spot volume ls REF NAME SIZE REGION ATTACHED vol-4b19c7 train-data 500 GB us-east i-7f3a2c $ spot volume rm vol-4b19c7 # must be detached; deletion is final
spot gpus#
$ spot gpus --min-vram 80 --tier spot SLUG NAME VRAM SPOT ONDEMAND AVAILABLE b200-sxm B200 SXM 180 GB $1.69 $2.49 16 h200-sxm H200 SXM 141 GB $0.75 $1.09 32 h200-nvl H200 NVL 141 GB $0.69 $0.99 16 h100-sxm H100 SXM 80 GB $0.55 $0.89 96
Scripting notes#
- Exit codes.
0success,1usage or validation error,2no capacity for that model and region,3insufficient balance,4authentication failure. --jsoneverywhere. Every command accepts it and prints one object or one array. Combine withjqrather than parsing tables.- Idempotence.
spot launch --idempotency-key <uuid>returns the same instance if the call is retried, which matters when a CI step reruns after a network error. - No hidden state. The CLI keeps nothing but the key: two machines with the same key see exactly the same objects.