Programmable AI cloud
Dashboard

Programmable cloud for AI workloads

Run AI workloads as resources.

Move from a local script to GPU compute, Jobs, Services, Sandboxes, Training, Inference, or RL without rebuilding the control plane for each step.

Open dashboard Read the resource model
One workload, one durable record
resource/job_01RUNNING
desired_state
running
observed_state
running
placement
H100 · us-east · on-demand
provider price
source rate attached
operation
op_01 · succeeded
usage
metered to project wallet
winterr run --gpu H100 python train.py
Transparent GPU supplyDeclarative resourcesVersioned workloadsOne wallet and API

The same system from compute to outcome.

Each product uses the same identity, resource, operation, billing, audit, and observability contracts. Start low in the stack and move up without changing clouds.

ComputeSearch normalized offers, lock a quote, and create Nodes or Clusters with the provider, region, topology, and source price recorded on the resource.
JobsRun containerized AI tasks with declared resources, secrets, volumes, timeouts, retries, logs, and deterministic cleanup.
ServicesDeploy HTTP or gRPC services with stable endpoints, health checks, replica targets, rollout state, logs, and provider-backed compute.
SandboxesCreate ephemeral interactive environments with terminal, file, process, network-policy, snapshot, and cleanup APIs behind a reviewed isolation boundary.
TemplatesVersion images, resources, storage, inputs, outputs, placement constraints, health checks, cost expectations, and benchmark claims as one immutable release.
BenchmarksRun versioned benchmark and evaluation matrices across models, engines, GPUs, providers, Templates, and configurations.
TrainingCompose versioned datasets, selected model revisions, training Templates, provider-backed compute, checkpoints, evaluations, and deployable artifacts.
InferenceDeploy selected open models behind dedicated endpoints with immutable engine configuration, health gates, token metering, and transparent provider compute.
RLCoordinate rollout inference, sandboxed environments, reward and verifier Jobs, trajectory storage, trainer Clusters, evaluations, and checkpoints.

Public API outside. Durable intent inside.

Every client reaches the same resource contract. The control plane records intent before a provider adapter changes infrastructure, then reconciles observed state until the operation is complete.

Review trust boundaries
  1. 01Client

    Python · TypeScript · CLI · MCP · dashboard

  2. 02Control plane

    Auth · validation · placement · wallet hold

  3. 03Provider adapter

    Capability-aware create and lifecycle operations

  4. 04Resource record

    Observed state · logs · usage · audit · cleanup

Start with one bounded Job.

Create a project, run a real workload, and follow its state, logs, usage, and cleanup from one identifier.

Open dashboard Review pricing