Read less.
Run more.

01Quickstart

Create an account; it takes an email address. Then, in the console:

1. Add credit on Billing. Koi is prepaid, and a launch needs enough for an hour of everything you’d be running.
2. Add an SSH key on Keys, so you can log in.
3. Launch on Machines: pick a GPU, how many, and optionally a region and image.
4. Log in. It takes minutes to start, and shows as Starting until SSH answers. Then the page gives you the command.

ssh koi@203.0.113.7
nvidia-smi

The user is always koi. Some machines answer SSH on another port, and the command includes it when they do. With VS Code and its Remote - SSH extension, “Open in VS Code” next to the command connects straight to the machine.

Terminate it when you’re done. Billing stops at once, and the machine and its disk are deleted, so copy off anything you want to keep first. An account can run 8 GPUs at once to start with. Write to us to raise that.

02From the API

Create an API key on the Keys page. It’s shown once. You’ll need an SSH key on your account too, which you can add there or with POST /ssh-keys.

export KOI_API_KEY=koi_...

# What can be launched right now, and for how much
curl https://console.koicompute.co/api/v1/skus \
  -H "Authorization: Bearer $KOI_API_KEY"

# Start one
curl https://console.koicompute.co/api/v1/instances \
  -H "Authorization: Bearer $KOI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"sku": "h100-1x", "name": "first-run", "stop_after_minutes": 120}'

That returns the machine with an id. Poll it until state is running and ssh is set:

curl https://console.koicompute.co/api/v1/instances/$ID \
  -H "Authorization: Bearer $KOI_API_KEY"

# And when you're done, which deletes the machine and its disk
curl -X DELETE https://console.koicompute.co/api/v1/instances/$ID \
  -H "Authorization: Bearer $KOI_API_KEY"

03CLI

One Python file that needs nothing but Python 3.8 or newer. Every command takes --json, and --help lists the rest, like koi spawn --drift and --container.

curl -fsSL https://console.koicompute.co/koi.py -o ~/.local/bin/koi && chmod +x ~/.local/bin/koi
koi login
koi keys add
koi spawn h100 --name first-run --ttl 2h
# ssh koi@203.0.113.7
koi release first-run

04Drift

Drift is a machine on spare capacity, at about half the price. When the capacity is needed back, your machine is reclaimed at once, without notice, and billing stops at that moment. The machine goes with its disk, so write checkpoints often to somewhere that outlives it. A Drift machine reclaimed before it’s ready costs nothing.

On machines that support it, resume_after_reclaim keeps the disk instead. The machine shows as paused, with the GPU unbilled and the disk still billed, until capacity returns and its container restarts. If that takes more than a day, the machine and disk are deleted.

curl https://console.koicompute.co/api/v1/instances \
  -H "Authorization: Bearer $KOI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"sku": "h100-8x", "spot": true, "container": "ghcr.io/acme/trainer:v3", "resume_after_reclaim": true}'

05Containers

Give a machine a public container image and it starts after boot, with GPU access. state says running before the container is up, and /var/log/koi-container.log on the machine shows how it’s going. Images without GPU drivers can’t run a container.

Pass secrets like a Hugging Face token as environment variables, with env on the API, --env NAME=value on the CLI (or just --env NAME to take it from your shell), or the box on the launch form. Koi hands the values to the machine and keeps only their names, so launching again asks for them again.

With stop_when_done, the machine and its disk are deleted when the container exits, so a finished job stops costing money, and we email you how it ended and what it cost. The instance keeps exit_code and the end of the output in container_log. An exit code of -1 means the container never started.

06Disks

Each machine comes with a disk, of the size in disk_gb, deleted with it. Disks are billed by the second, at disk_price_per_hour, until they’re deleted.

A standard disk ($0.08 per GB-month) has no redundancy, so its data is lost if the hardware fails. A redundant disk ($0.15 per GB-month) is mirrored to three drives, and isn’t on every machine. Neither outlives the machine.

07Billing

You pay by the second, from when the provider has your machine running until you ask it to stop or it’s reclaimed. A launch that never becomes ready costs nothing. A running machine’s price doesn’t change, whatever happens to prices after you launch it.

We email you when what’s running has under two hours of credit left. When your balance reaches zero, every machine stops and its disk is deleted, and we email you to say so.

Automatic top-ups charge a saved card when your credit falls below a level you choose while something’s running, up to five times in 24 hours, so a missed email can’t stop your machines. Turn them on under Billing.

Stop after ends a machine by itself after a time you set, so a forgotten one can only cost so much. Its disk goes with it.

Spending limits cap what a key’s machines may cost, in all, for handing a key to a script or an agent. A launch needs an hour under the limit, and machines stop when they reach it. Set one on the Keys page. A key can’t raise its own.

08API reference

The base URL is https://console.koicompute.co/api/v1. Send your key as Authorization: Bearer koi_…. Money is in dollars, as strings, like "1.990000". An error is { "error": { "message": "…" } }, with a message you can show to a person.

Endpoints

Launching: POST /instances

Machine states

Errors

The API calls machines instances, in its paths and fields. A SKU’s vcpus and memory_gb are the least you’d get. A machine’s say what it really has.

09Not built yet

Shoal (clusters), Current (serverless endpoints), Pond (storage that outlasts a machine) and a Python SDK aren’t available yet, so they aren’t documented here.