AfterQuery Compute TypeScript SDK

Node.js 18+ client for the AfterQuery Compute API.

import { Compute, GpuCreateError } from '@afterquery/compute'

const compute = new Compute({ apiKey: '...' })

try {
  const sandbox = await compute.create(
    {
      image: 'registry.example.com/team/image:tag',
      resources: { gpu: { type: 'l4', count: 1 } },
    },
    { idempotencyKey: 'job-42-attempt-1' },
  )
} catch (error) {
  if (error instanceof GpuCreateError) {
    // Preserve these values when reconciling an ambiguous create outcome.
    console.error(error.idempotencyKey, error.sandboxId)
  }
  throw error
}

GPU creates require a stable idempotency key. If one is not supplied, the SDK generates one and exposes it through GpuCreateError on any create or readiness failure. GPU batch creation is intentionally unsupported.

For an ambiguous failure without a sandbox ID, call cancelGpuCreate(error.idempotencyKey) to request cleanup using the stable key. The SDK retries one transport-ambiguous cancellation request with the same key. Deleted and Failed are immutable terminal history; CleanupPending means cleanup is still converging, so call cancelGpuCreate again until a terminal state is returned. A same-key create retry may also return a terminal handle without starting or waiting for a new sandbox.

See the repository for API setup and the complete SDK surface.

Cleaning up from inside a sandbox

Code running inside a sandbox can delete its own sandbox as its final act, so a finished one-shot job never lingers (and Kubernetes never restarts it):

import { Compute } from '@afterquery/compute'

// ... do the work ...
await new Compute().selfDelete()   // remove me; nothing left running

No API key is needed: the platform injects COMPUTE_SELF_TOKEN and COMPUTE_API_URL into every sandbox, and that token can only delete the sandbox it belongs to. Outside a sandbox the call throws.