Skip to content
OpenShell

Set Up OpenShell on Kubernetes

Deploy the OpenShell gateway to a Kubernetes cluster using the official Helm chart from GHCR.

OpenShell on Kubernetes: the gateway provisions a Sandbox resource, separate supervisor and workload Pods, a boundary Service, NetworkPolicy, and a workspace PVC.

The gateway’s Kubernetes compute driver creates a Sandbox resource and a separate supervisor Pod. The Agent Sandbox controller creates the workload Pod from that resource. A Service gives the supervisor a stable address for the workload’s authenticated TLS boundary, and a PVC preserves /sandbox across Pod restarts. The gateway can run in the same namespace as these resources or in a separate namespace.

The supervisor opens the boundary connection and maintains an outbound session to the gateway. A NetworkPolicy allows only supervisor ingress to the workload’s boundary port and denies new workload-initiated connections. Agent network requests return on the established channel; the supervisor evaluates policy and opens approved upstream connections. For the trust boundaries and request flow, see Architecture.

Make sure the following are in place before you install.

Prerequisite Required Notes
Kubernetes 1.29+ with RBAC enabled Yes No additional notes.
CNI that enforces ingress and egress NetworkPolicy in sandbox namespaces Yes Verify enforcement on your cluster.
Helm 3.x Yes No additional notes.
Agent Sandbox controller and CRDs Yes Install before the OpenShell chart. Refer to Install Agent Sandbox.
cert-manager No Refer to Managing Certificates. Use cert-manager only if you prefer it over the built-in PKI job.
Kubernetes Gateway API No Refer to Ingress. Use it only for external access without port-forwarding.

OpenShell uses the Agent Sandbox Kubernetes SIG project to provision sandbox pods. Install the Agent Sandbox controller and its CRDs on your cluster before installing the OpenShell Helm chart.

Apply the latest release manifest:

Terminal window
kubectl apply -f https://github.com/kubernetes-sigs/agent-sandbox/releases/latest/download/sandbox.yaml

This creates the agent-sandbox-system namespace, installs the sandboxes.agents.x-k8s.io CRD, and starts the controller.

The Helm chart checks for a supported Agent Sandbox API before it creates gateway resources. This preflight is enabled by default. Disable it only for offline helm template rendering, where Helm cannot discover cluster APIs:

Terminal window
helm template openshell oci://ghcr.io/nvidia/openshell/helm-chart \
--version <version> \
--set agentSandbox.preflight.enabled=false

The chart does not install or upgrade the cluster-scoped Agent Sandbox CRDs or controller.

Confirm the controller pod is running before proceeding:

Terminal window
kubectl -n agent-sandbox-system get pods

The controller pod should reach Running status within a few seconds. For cluster-specific setup instructions, including KinD and GKE walkthroughs, refer to the Agent Sandbox getting started guide.

OpenShell detects the served Agent Sandbox Sandbox API when the Kubernetes gateway first needs it and caches that choice for the gateway process. If you upgrade Agent Sandbox in place, restart the OpenShell gateway after the Agent Sandbox controller and CRD rollout completes so the gateway can detect the served API versions again. Existing sandboxes keep running during the upgrade, and the restarted gateway can continue managing them.

Terminal window
kubectl create namespace openshell

Install from the OCI registry on GHCR. Replace <version> with the chart version you want to install.

Terminal window
helm upgrade --install openshell \
oci://ghcr.io/nvidia/openshell/helm-chart \
--version <version> \
--namespace openshell

To use the latest development build instead of a stable release:

Terminal window
helm upgrade --install openshell \
oci://ghcr.io/nvidia/openshell/helm-chart \
--version 0.0.0-dev \
--namespace openshell

The chart automatically generates PKI secrets on first install using pre-install Helm hooks. No manual secret creation is required.

To add labels or annotations to both certificate hook Pod templates, use pkiInitJob.podLabels and pkiInitJob.podAnnotations. Chart-owned Pod labels take precedence when custom keys conflict. For example, disable sidecar injection for hook Pods in an Istio-injected namespace:

pkiInitJob:
podAnnotations:
sidecar.istio.io/inject: "false"

The default Kubernetes driver puts all workspaces in one sandbox namespace. To have the gateway create a namespace for each workspace, follow Workspace on Kubernetes.

For a platform-managed namespace, install the gateway without namespace-scoped sandbox resources, then install the workspace chart in the sandbox namespace:

Terminal window
helm upgrade --install openshell \
oci://ghcr.io/nvidia/openshell/helm-chart \
--version <version> \
--namespace openshell \
--set workspaceResources.enabled=false \
--set server.sandboxNamespace=app-a
helm upgrade --install openshell-workspace \
oci://ghcr.io/nvidia/openshell/openshell-workspace \
--version <version> \
--namespace app-a \
--set gateway.serviceAccount.name=openshell \
--set gateway.serviceAccount.namespace=openshell

The workspace chart does not create the namespace or deploy a gateway. It owns only the sandbox ServiceAccount, Role, RoleBinding, and NetworkPolicy in its release namespace. For one pre-provisioned namespace, keep server.drivers.kubernetes.workspaceMode=shared and set server.sandboxNamespace=app-a. To map multiple workspaces to separately provisioned namespaces, use workspaceMode=operator, configure exactly one of operatorNamespaceLabel or operatorNamespaceFile, and install the workspace chart in every allowlisted namespace. In operator mode, the workspace chart Role is the gateway’s only Secret permission grant in that namespace.

If you enable server.drivers.kubernetes.allowDriverConfig on the gateway, also set gateway.allowDriverConfig=true on every workspace release. This grants the namespace-scoped PVC metadata read needed to admit caller-selected PVC mounts. Leave both values disabled when callers do not need driver config.

Store provider credentials in Kubernetes Secrets

Section titled “Store provider credentials in Kubernetes Secrets”

To store provider credentials as Kubernetes Secrets instead of in the gateway database, enable the Kubernetes Secrets credential driver with a dedicated namespace:

Terminal window
helm upgrade --install openshell \
oci://ghcr.io/nvidia/openshell/helm-chart \
--version <version> \
--namespace openshell \
--set server.credentialDrivers.kubernetesSecrets.enabled=true \
--set server.credentialDrivers.kubernetesSecrets.namespace=openshell-credentials \
--set server.credentialDrivers.kubernetesSecrets.createNamespace=true

The driver stores every credential in that namespace, in every workspace mode. The gateway receives Secret permissions through a Role in that namespace only. For the full set of options, refer to the gateway configuration reference.

Terminal window
kubectl -n openshell rollout status statefulset/openshell

If you set workload.kind=deployment, wait on the Deployment instead:

Terminal window
kubectl -n openshell rollout status deployment/openshell

For local evaluation, use a port-forward:

Terminal window
kubectl -n openshell port-forward svc/openshell 8080:8080

The chart generates an mTLS bundle for transport security. Kubernetes deployments do not use that bundle as user authentication; configure OIDC or a trusted access proxy as described in Access Control. For local port-forwarded access, copy the generated bundle so the CLI can verify the gateway certificate:

Terminal window
mkdir -p ~/.config/openshell/gateways/k8s/mtls
kubectl -n openshell get secret openshell-client-tls \
-o jsonpath='{.data.ca\.crt}' | base64 -d > ~/.config/openshell/gateways/k8s/mtls/ca.crt
kubectl -n openshell get secret openshell-client-tls \
-o jsonpath='{.data.tls\.crt}' | base64 -d > ~/.config/openshell/gateways/k8s/mtls/tls.crt
kubectl -n openshell get secret openshell-client-tls \
-o jsonpath='{.data.tls\.key}' | base64 -d > ~/.config/openshell/gateways/k8s/mtls/tls.key

The server certificate SANs include localhost and 127.0.0.1, so hostname verification passes over the port-forward without extra flags.

In another terminal, register the gateway with the user authentication mode you configured and verify it is reachable. For example, with OIDC:

Terminal window
openshell gateway add https://127.0.0.1:8080 --local --name k8s \
--oidc-issuer https://your-idp.example.com/realms/openshell \
--oidc-client-id openshell-cli
openshell status

To upgrade, rerun helm upgrade --install with the new --version and the values you installed with. Running sandboxes keep their supervisor and sandbox runtime images until they stop. When a stopped sandbox starts, it takes both images from the upgraded gateway and keeps its workload image and workspace.

If you override images, set supervisor.image and sandboxRuntime.image to the same release, or set only global.image.tag. An incompatible pair fails when a sandbox starts, and the supervisor log shows the cause while the start is in progress:

Terminal window
kubectl -n openshell logs -l openshell.ai/sandbox-id=<sandbox-id>,openshell.ai/component=supervisor

After you correct the images, recover a sandbox whose start timed out as described in Sandbox Lifecycle.

The Helm chart renders gatewayConfig into the mounted gateway.toml ConfigMap. Each top-level key names a TOML table, and its fields use the names in the gateway configuration reference. For example, set the gateway log level with:

Terminal window
helm upgrade openshell \
oci://ghcr.io/nvidia/openshell/helm-chart \
--version <version> \
--namespace openshell \
--set 'gatewayConfig.openshell\.gateway.log_level=debug'

Keep credentials, database URLs, and private keys in Kubernetes Secrets, using the chart’s environment-variable, file, or volume wiring. The chart rejects database_url, inline URL credentials, and PEM private keys in gatewayConfig, but it does not detect every possible secret.

Helm values for Services, probes, TLS certificates, Secret references, RBAC, NetworkPolicies, and mounts still control those Kubernetes resources. The chart derives their corresponding runtime settings to keep the gateway configuration aligned with the deployed resources.

YAML scalar values retain their types, scalar lists become TOML arrays, and nested maps become inline tables. Top-level lists become arrays of tables and preserve item order. Null fields are omitted; null array elements are invalid. Helm template expressions in strings are not evaluated.

Existing application values remain supported as deprecated aliases. Explicit gatewayConfig fields take precedence, including resource-admission settings and the Kubernetes Secrets credential namespace. The global image pull policy applies to runtime and supervisor images. Leave image_pull_policy and sandbox.image.pullPolicy unset to use Kubernetes defaults for the workload image.

The most commonly changed values are:

Value Purpose
global.image.registry / global.image.tag / global.image.pullPolicy Shared registry, tag, and pull policy for the gateway, supervisor, and trusted sandbox runtime. Individual image settings take precedence.
gateway.image.registry / gateway.image.repository / gateway.image.tag / gateway.image.digest Gateway container image. Defaults to ghcr.io/nvidia/openshell/gateway:latest.
replicaCount Number of gateway replicas. Values above 1 require shared PostgreSQL through server.externalDbSecret.
priorityClassName Existing Kubernetes PriorityClass for gateway pods. Empty by default.
topologySpreadConstraints Kubernetes topology spread constraints for gateway replicas. Empty by default. Set selectors to match the gateway pod labels.
workload.kind Gateway workload controller. Use statefulset for SQLite or deployment with server.externalDbSecret.
workload.allowMultiReplicaStatefulSet Allow replicaCount > 1 with workload.kind=statefulset. Prefer Deployment for external database-backed multi-replica gateways.
autoscaling.enabled / autoscaling.minReplicas / autoscaling.maxReplicas Render a HorizontalPodAutoscaler instead of a fixed replicaCount. Refer to High Availability.
server.sandboxNamespace Namespace where sandbox pods are created. Defaults to the Helm release namespace when left empty.
workspaceResources.enabled Create namespace-scoped sandbox prerequisites from the gateway chart. Disable when installing the workspace chart separately.
server.externalDbSecret Secret containing a PostgreSQL connection URI in the uri key. Use when the database is managed outside the chart.
server.dataStorageSize Storage request for the gateway data PVC under workload.kind=statefulset. Defaults to 1Gi. Requires an explicit unit, such as 20Gi.
server.dataStorageClass StorageClass for the gateway data PVC under workload.kind=statefulset. Empty (default) omits storageClassName and uses the cluster’s default StorageClass. Use a block-backed class, not NFS.
server.telemetryEnabled Enable anonymous OpenShell telemetry from the gateway and its sandbox supervisors. Set to false to opt out.
sandbox.image.repository / sandbox.image.tag / sandbox.image.digest Default sandbox image used when a sandbox does not specify one.
sandboxRuntime.image.registry / sandboxRuntime.image.repository / sandboxRuntime.image.tag / sandboxRuntime.image.digest Trusted workload-side image that provides the openshell-sandbox binary. A digest takes precedence over the tag.
supervisor.image.registry / supervisor.image.repository / supervisor.image.tag / supervisor.image.digest Trusted control-side supervisor image.
server.sandboxImagePullSecrets Image pull secrets attached to sandbox pods. Referenced Secrets must exist in the sandbox namespace.
server.grpcEndpoint Endpoint that sandbox supervisors use to call back to the gateway. Must be reachable from inside the cluster.
server.disableTls Run the gateway over plaintext HTTP. Use only behind a trusted transport.
server.auth.allowUnauthenticatedUsers Accept user-facing calls without OIDC or mTLS credentials. Use only for trusted local development or a fully trusted access proxy.
server.enableLoopbackServiceHttp Enable local plaintext HTTP for loopback sandbox service URLs. Defaults to true.
server.serviceHostnameFormat Set to hashed to make exposed-service hostnames opaque per-gateway hash labels when multiple gateways share one service base domain behind a single ingress; the hash discriminator is the gateway installation name (server.name, defaulting to the release name), which must be unique among those gateways. Empty keeps the readable composite hostnames. Refer to Ingress.
pkiInitJob.serverDnsNames / certManager.serverDnsNames Additional gateway server DNS SANs. Wildcard SANs also enable sandbox service URLs under that domain.
supervisor.sandboxRuntime.boundaryPort Non-privileged TLS port used between paired supervisor and sandbox Pods.
upstreamProxy Operator-owned corporate HTTP forward proxy for policy-approved TLS egress. Refer to Configure a Corporate Upstream Proxy.

Use a values file for repeatable deployments:

Terminal window
helm upgrade --install openshell \
oci://ghcr.io/nvidia/openshell/helm-chart \
--version <version> \
--namespace openshell \
--values my-values.yaml

To use private sandbox images, create a kubernetes.io/dockerconfigjson Secret in the sandbox namespace and reference its name:

Terminal window
kubectl -n openshell create secret docker-registry regcred \
--docker-server=registry.example.com \
--docker-username="$REGISTRY_USER" \
--docker-password="$REGISTRY_TOKEN"
server:
sandboxImage: registry.example.com/team/openshell-sandbox:latest
sandboxImagePullSecrets:
- name: regcred

Configure a corporate forward proxy when sandbox TLS egress cannot dial the Internet directly. OpenShell evaluates policy and SSRF checks before it opens an HTTP CONNECT tunnel through the proxy. The proxy URL is operator-owned configuration. Sandbox environment variables cannot select, replace, or bypass it.

Create the credential Secret in the sandbox namespace when the proxy requires Basic authentication. The Secret value uses the user:pass form.

Terminal window
kubectl -n openshell create secret generic corporate-proxy-auth \
--from-literal=credentials="$PROXY_USER:$PROXY_PASSWORD"

Add the proxy settings to your Helm values file. Replace the DNS suffixes and CIDRs in noProxy with values for your cluster. noProxy bypasses only the corporate proxy. OpenShell policy evaluation still applies.

upstreamProxy:
url: http://proxy.corp.example:8080
noProxy: .svc,.svc.cluster.local,10.96.0.0/12,10.244.0.0/16
authSecret:
name: corporate-proxy-auth
key: credentials
authAllowInsecure: true

Use authAllowInsecure: true only when you accept that Basic authentication is cleartext on the connection to an http:// proxy. The initial release supports http:// proxy endpoints and TLS CONNECT egress. It does not support HTTPS-to-proxy, custom corporate CA bundles, or forwarding plain HTTP egress through the proxy.

The credential mounts only in the separately scheduled supervisor Pod. The sandbox workload cannot read it through its environment or volumes.

The chart creates the following RBAC resources in the release namespace:

Resource Scope Name
ServiceAccount Namespace openshell
ServiceAccount Namespace openshell-sandbox (for sandbox pods)
Role + RoleBinding Namespace openshell-sandbox
ClusterRole + ClusterRoleBinding Cluster openshell-node-reader-<release namespace>

Every other object the chart creates is namespaced. The ClusterRole and ClusterRoleBinding in the last row are the only cluster-scoped objects.

When the Kubernetes Secrets credential driver is enabled, the chart also creates an openshell-credential-secrets Role and RoleBinding in the credential namespace. That Role grants get, create, patch, and delete on all Secrets in the credential namespace, and is the gateway’s only permission grant for provider credential Secrets.

In managed and operator workspace modes, the chart creates an openshell-workspace-secret-source Role and RoleBinding in the sandbox namespace. That Role grants get on only the gateway client TLS Secret and, in managed mode, the configured image-pull Secrets. The gateway stages their contents into each sandbox runtime generation’s Secrets in the workspace namespace.

The namespaced Role covers sandbox lifecycle, supervisor resources, and identity. Both the gateway chart and separate workspace chart grant these permissions. Stopped-sandbox updates use direct Pod get requests to check known workload and supervisor Pod names for remaining Pods.

API Group Resource Verbs
agents.x-k8s.io sandboxes, sandboxes/status create, delete, get, list, patch, update, watch
"" events get, list, watch
"" pods create, delete, get, list, patch, watch
"" services create, get
"" secrets create, delete
networking.k8s.io networkpolicies create, get

The ClusterRole grants node inspection and token validation:

API Group Resource Verbs
authentication.k8s.io tokenreviews create
"" nodes get, list, watch

To use an existing ServiceAccount instead of creating one, set serviceAccount.create=false and supply its name:

Terminal window
helm upgrade --install openshell \
oci://ghcr.io/nvidia/openshell/helm-chart \
--version <version> \
--namespace openshell \
--set serviceAccount.create=false \
--set serviceAccount.name=my-existing-sa

The ServiceAccount must already have the Role and ClusterRole bindings described above.

By default the release creates the ClusterRole and ClusterRoleBinding, so the installer needs cluster-scoped permissions. When cluster-scoped RBAC is owned by a different team, or the installer is a namespace-admin GitOps controller, split the install into a cluster-admin step and a namespace-admin step.

A cluster-admin applies the cluster-scoped objects once per gateway ServiceAccount. Render them from the same values the release uses, so the rules match the configured workspace mode and credential driver:

Terminal window
helm template openshell \
oci://ghcr.io/nvidia/openshell/helm-chart \
--version <version> \
--namespace openshell \
-f my-values.yaml \
--set rbac.create=true \
--set rbac.clusterScoped.create=true \
--set agentSandbox.preflight.enabled=false \
--show-only templates/clusterrole.yaml \
--show-only templates/clusterrolebinding.yaml | kubectl apply -f -

A namespace-admin then installs and upgrades the chart with cluster-scoped objects omitted. The release renders only namespaced objects and needs no permission on clusterroles or clusterrolebindings:

Terminal window
helm upgrade --install openshell \
oci://ghcr.io/nvidia/openshell/helm-chart \
--version <version> \
--namespace openshell \
-f my-values.yaml \
--set rbac.clusterScoped.create=false

The gateway ServiceAccount name and namespace are unchanged by this flag, so the pre-created ClusterRoleBinding still binds the ServiceAccount the namespaced release creates. The same holds with serviceAccount.create=false: the binding subject follows serviceAccount.name, so pass the same values to both steps.

Helm deletes objects that leave a release manifest, so setting rbac.clusterScoped.create=false on a release that already owns the ClusterRole and ClusterRoleBinding deletes them. The gateway then loses TokenReview until a cluster-admin re-applies them. Hand ownership over first, as cluster-admin, so nothing is deleted:

Terminal window
kubectl annotate clusterrole "openshell-node-reader-<namespace>" \
helm.sh/resource-policy=keep --overwrite
kubectl annotate clusterrolebinding "openshell-node-reader-<namespace>" \
helm.sh/resource-policy=keep --overwrite

The objects then survive the upgrade that sets the flag, and the cluster-admin owns them from that point on. Fresh installs need no such step.

rbac.clusterScoped.create is independent of the workspace mode. managed and operator modes change what the ClusterRole grants, but they never require a namespace-admin to apply it. Re-run the cluster-admin step after changing any value that affects the ClusterRole rules.

Set rbac.clusterScoped.clusterRoleName and rbac.clusterScoped.clusterRoleBindingName when the cluster-admin owns the naming.

In the default shared workspace mode the release also creates a namespaced sandbox Role granting Agent Sandbox (agents.x-k8s.io) permissions. Kubernetes forbids granting permissions you do not hold, and the built-in admin ClusterRole does not cover that CRD, so an installer holding only admin cannot create it. rbac.clusterScoped.create=false alone is then not enough, and the install fails with attempting to grant RBAC permissions not currently held.

Workspace mode Installer holds Use
shared built-in admin only rbac.create=false, cluster-admin pre-creates all gateway RBAC
shared admin plus the sandbox permissions in the namespace rbac.clusterScoped.create=false
managed, operator built-in admin only rbac.clusterScoped.create=false

managed and operator render no namespaced sandbox Role, so no extra grant is needed there.

With rbac.create=false, the cluster-admin applies the namespaced RBAC in the same step by adding it to the render:

Terminal window
helm template openshell \
oci://ghcr.io/nvidia/openshell/helm-chart \
--version <version> \
--namespace openshell \
-f my-values.yaml \
--set rbac.create=true \
--set rbac.clusterScoped.create=true \
--set agentSandbox.preflight.enabled=false \
--show-only templates/clusterrole.yaml \
--show-only templates/clusterrolebinding.yaml \
--show-only templates/role.yaml \
--show-only templates/rolebinding.yaml \
--show-only templates/peer-role.yaml | kubectl apply -f -

To grant the installer the sandbox permissions instead, bind it to a Role carrying the same rules as the chart’s openshell-sandbox Role.

The gateway workload and the PKI bootstrap hooks satisfy the Restricted Pod Security Standard, so the release installs into a namespace labeled pod-security.kubernetes.io/enforce=restricted.

The gateway exposes /healthz for process liveness and /readyz for dependency-aware readiness on the health port. The Helm chart wires both into Kubernetes probes:

  • startupProbe and livenessProbe use /healthz.
  • readinessProbe uses /readyz, which reflects the latest result of an in-process background database check.

A gateway that starts while its database is unreachable keeps retrying the connection. Meanwhile /healthz succeeds and /readyz reports not ready, so Kubernetes keeps the pod without routing traffic to it.

Each sandbox supervisor Pod uses a tcpSocket readiness probe on port 5501. The supervisor accepts connections on that port only while its gateway session is ready. If you add a default-deny ingress policy to sandbox namespaces, allow TCP port 5501 from the node so kubelet can reach supervisor Pods.

We are a Cloud Native Computing Foundation sandbox project.