Skip to content
Python
2.0 KB
Raw Blame

Deployment options

Start with the simplest placement that fits your product.

Choose a platform

Platform Reference
Private Kubernetes cluster Your cluster, PostgreSQL and OpenShell
AWS EKS, PostgreSQL, S3 and OpenShell
GCP GKE Standard, Cloud SQL, Cloud Storage and OpenShell
Azure AKS, PostgreSQL, Blob and OpenShell

The application-facing service API stays the same. Cloud identity, networking, storage and lifecycle setup are provider-specific. NVIDIA OpenShell is the sandbox integration across these reference deployments.

Choose execution placement

Placement Use it when
Managed local harness A desktop app or one container owns a workspace
Service with nearby execution Several API nodes need durable sessions and shared worker capacity
Service with separate execution Sandboxes need different hardware, capacity or lifecycle from the agent workers

SQLite serves local applications. PostgreSQL supports service deployments and can also be selected through the Rust SDK. Database choice does not remove team/session capabilities. Bind an execution environment explicitly when agents must share it; sharing does not resolve conflicting file edits or ports automatically.

OpenShell runs tool execution in a sandbox while the current harness remains in the trusted service domain. RuntimeProfiles select the provider and allowed resources. A profile is configuration; a provisioned environment is an instance.

Start with the local service setup and service CLI. Cloud credentials, budgets and cluster identity are operator-owned. Cloud deployment paths are experimental; validate the configuration in the target environment before production use.

Before upgrading, read versions and compatibility and the storage migration instructions.