Kovara writes SkyPilot task files whose candidate list is ranked from Kovara's own market data. SkyPilot then launches the task on your own provider accounts and fails over down the list until one candidate starts. Kovara launches nothing, never sees your provider credentials and bills nothing for a task file.
Three ways to get one:
- Website: Deployments → SkyPilot tasks (
/deployments/skypilot, signed in). - CLI:
kovara skypilot task --gpu GPU0015 --gpus 8(see the CLI reference). - API:
GET /api/public/v1/skypilot?gpu=GPU0015&gpus=8&format=yaml(no key), ormarket.skypilotYaml()in the TypeScript SDK.
How the order is decided
resources.ordered is written in this order:
- Capacity reported now. For every candidate, Kovara looks at the provider's latest capacity reading for that GPU and GPU count. A reading counts only when it is at most 15 minutes old, the same rule the rest of Kovara uses for live availability. Each site that reported Available becomes one candidate pinned to that region (
infra: lambda/us-east-1). Sites that reported Limited follow. - Fallbacks. Clouds with no positive reading come next, without a region, so SkyPilot searches the regions its own catalog lists for that instance type. Clouds whose latest reading was Unavailable everywhere come last. Turn fallbacks off (
fallbacks=false,--no-fallbacks) to write only places with capacity reported now. - Within each group, by Kovara's listed price per GPU-hour, cheapest first. Places without a comparable USD price come after priced ones.
The file's header comments repeat the order with the reason for each place, the reading time and the listed price with its offering ID. The prices are listings, not quotes.
What is exact, and what is left out
| Fact | Where it comes from | Rule |
|---|---|---|
| Instance type | The provider's own API as Kovara recorded it (lib/byok/adapters/lambda-catalogue.ts, verda-catalogue.ts) | Used only if SkyPilot's catalog also lists the type with the same GPU count. Otherwise sky launch would refuse it, so it is never written. |
| Region | The site code of a fresh capacity reading | Pinned only when SkyPilot's catalog lists that region (or zone) for the type. Anything else is left to SkyPilot. |
| Capacity | live_availability_public_readings(): the latest run per provider, at most 15 minutes old | Same provider, GPU and stated GPU count. Older or future-dated readings are never used. |
| Price | Current public offerings | Same provider, GPU, GPUs per node and contract basis (On-Demand or Spot) only. Never converted: a price in another currency is shown as listed and is not ranked. |
| Spot | SkyPilot's catalog | Spot is written only for types SkyPilot's catalog prices as spot. Kovara does not read spot capacity, so spot candidates are never pinned to a region. |
Clouds today are Lambda and Verda. These are the providers whose instance-type names Kovara holds from their own APIs and that the SkyPilot client supports. Scaleway's types are recorded too, but SkyPilot has no Scaleway support, so they are not written. More clouds are added as Kovara records more providers' APIs.
The SkyPilot catalog snapshot (lib/skypilot/catalog-snapshot.ts) names the catalog version and commit it was taken from. To refresh it:
git clone --depth 1 https://github.com/skypilot-org/skypilot-catalog.git /tmp/skypilot-catalog
python3 -I scripts/skypilot/snapshot-catalog.py /tmp/skypilot-catalog > lib/skypilot/catalog-snapshot.ts
NODE_PATH=$(npm root -g) node scripts/check-skypilot.cjsThe task file
# SkyPilot task generated by Kovara (kovaragrid.com) at 2026-11-11T15:09:13.011Z
# NVIDIA A100 40GB SXM x8 per node, on-demand. SkyPilot tries resources.ordered from the top:
# 1. Lambda lambda/us-east-1 gpu_8x_a100 (A100:8): capacity reported at 2026-11-11T15:00:01Z; $1.99/GPU-hour listed (OFR-0232)
# ...
name: "a100-run"
resources:
use_spot: false
ordered:
- infra: "lambda/us-east-1"
instance_type: "gpu_8x_a100"
secrets:
"HF_TOKEN": null
run: |
python train.py- Every string is double-quoted, keys included. SkyPilot reads YAML 1.1, where a bare
YES,NO,ONorOFFwould become a boolean. - Secrets are declared by name with no value. SkyPilot then requires each one at launch:
sky launch … --secret HF_TOKENreads it from your shell. Environment variables (envs) are written into the file, so do not put tokens there. - Each ordered entry names only the instance type. The type fixes the accelerators; the header comment shows them.
setupandrunare kept exactly as given. Control characters other than tab and newline are refused.
Launching
pip install "skypilot[lambda]" # Verda needs no extra
sky check lambda verda # after adding your provider credentials as SkyPilot's docs describe
sky launch -c a100-run a100-run.sky.yaml --secret HF_TOKEN -i 30 --down
sky status a100-run
sky down a100-run-i 30 --down tears the cluster down after 30 idle minutes. Autostop is on by default in the builder, the CLI and the API; turn it off with autostop=off. Billing is between you and the provider.
API reference
GET /api/public/v1/skypilot without gpu lists every GPU and GPU count a task can be written for, with each cloud's instance type, whether capacity is reported now and the listed on-demand price.
GET /api/public/v1/skypilot?gpu=…&gpus=… returns the ranking, what was left out and why, the file and the launch commands.
| Parameter | Default | Meaning |
|---|---|---|
gpu | required | Kovara GPU ID, for example GPU0015 |
gpus | required | GPUs per node |
spot | false | Spot instead of on-demand |
clouds | all | Comma-separated: lambda, verda |
fallbacks | true | Keep clouds without reported capacity as fallbacks |
nodes, name, workdir, setup, run | 1, kovara-task, none, none, nvidia-smi | The workload |
env | none | NAME=value; repeat the parameter for several |
secrets | none | Comma-separated names |
disk_size, ports | SkyPilot's defaults | GB; comma-separated ports or ranges |
autostop | 30 | Idle minutes, or off; used in the launch command |
format | json | yaml returns the file alone (application/yaml) |
Responses are cached for 60 seconds, because capacity readings stay current for only 15 minutes.