Install the CASA Runtime

The CASA runtime is distributed as the casa-runtime Helm chart, published to GHCR as an OCI artifact. It bundles both the control-plane components and the sidecar via the ext-auth-service subchart.

Install from GHCR (recommended)

First, create the namespace and enable Istio sidecar injection:

kubectl create namespace casa-dev
kubectl label namespace casa-dev istio-injection=enabled

Without the istio-injection=enabled label, agent and MCP pods will not receive Istio sidecar proxies and CASA enforcement will not work.

Then install the chart.

At minimum you must supply the LLM endpoint credentials — everything else has a working default:

helm install casa-dev oci://ghcr.io/outshift-open/helm/casa-runtime \
  --version 0.1.17 \
  --namespace casa-dev \
  --set authService.openai.apiBaseUrl="https://your-llm-endpoint" \
  --set authService.openai.jwtToken="YOUR_PIPELINE_TOKEN" \
  --set authService.openai.llmApiBaseUrl="https://your-llm-endpoint" \
  --set authService.openai.llmApiKey="YOUR_LLM_KEY" \
  --set authService.openai.modelId="gpt-4o" \
  --set authService.openai.pipelineModelId="gpt-4o"

No helm repo add is needed — OCI charts are pulled directly. All images are public; no registry authentication is required.

Default passwords authService.database.password and authService.idp.adminPassword default to postgres/admin. Override them for any non-local deployment — see Production Configuration.

Install from source

# Clone the repo, then create the namespace:
kubectl create namespace casa-dev
kubectl label namespace casa-dev istio-injection=enabled

helm dependency build deployments/helm/casa-runtime/
helm install casa-dev deployments/helm/casa-runtime/ \
  --namespace casa-dev

helm dependency build is required to download the sidecar subchart before the first install from source.

What gets installed

The chart deploys the full control plane plus the sidecar in a single release:

Component	Description
`casa-dev-auth-service`	FastAPI authorization server
`casa-dev-ui-explorer`	CASA Explorer web UI
`casa-dev-keycloak`	Keycloak IdP (pre-configured)
`casa-dev-postgres-auth`	PostgreSQL for the auth service
`casa-dev-postgres-keycloak`	PostgreSQL for Keycloak
`casa-dev-operator`	Kubernetes operator (reconciles `MultiAgentSystem` CRDs)
`casa-dev-sidecar-*`	Ext-auth service, otel-collector, Jaeger, eBPF instrumentation

Wait for all pods to be ready:

kubectl -n casa-dev wait --for=condition=ready pod --all --timeout=300s

Production Configuration

For production deployments, create a values-prod.yaml that configures credentials, the LLM backend, and ingress.

Credentials & LLM backend

CASA requires credentials for the PostgreSQL database, the Keycloak admin account, and an OpenAI-compatible LLM endpoint. The auth service uses two OpenAI client instances internally — both for the authorization pipeline, not for the demo agents:

apiBaseUrl / jwtToken — used by the LLM verifier and hybrid tool matcher (bearer-token auth)
llmApiBaseUrl / llmApiKey — used by the task extractor and tool matcher (API-key auth)

In practice both pairs point to the same endpoint. They exist as separate config because the pipeline components were built with different OpenAI client setups:

authService:
  database:
    password: "CHANGE_ME"
  idp:
    adminPassword: "CHANGE_ME"
  openai:
    apiBaseUrl: "https://your-litellm-or-openai-endpoint"     # pipeline LLM verifier
    jwtToken: "CHANGE_ME"                                      # bearer token for above
    llmApiBaseUrl: "https://your-litellm-or-openai-endpoint"  # pipeline task extractor
    llmApiKey: "CHANGE_ME"                                     # API key for above
    modelId: "bedrock/global.anthropic.claude-sonnet-4-6"     # model for both pipeline clients
    pipelineModelId: "azure/gpt-4o"                           # model for embedding/tool matching

All six openai.* fields are required — the auth service will start but authorization checks will fail silently if any are empty.

Ingress

Each component exposes an ingress block. Set className, apiDomainName, and domainPrefix — the resulting hostname is <domainPrefix>.<apiDomainName>. Use annotations for cert-manager or any other ingress annotation:

authService:
  ingress:
    enabled: true
    className: "nginx-internal"
    apiDomainName: "dev.eticloud.io"
    domainPrefix: "casa-auth"        # → casa-auth.dev.eticloud.io
    annotations:
      cert-manager.io/cluster-issuer: letsencrypt

uiExplorer:
  ingress:
    enabled: true
    className: "nginx-internal"
    apiDomainName: "dev.eticloud.io"
    domainPrefix: "casa-explorer"    # → casa-explorer.dev.eticloud.io
    annotations:
      cert-manager.io/cluster-issuer: letsencrypt

keycloak:
  hostname: "casa-keycloak.dev.eticloud.io"
  hostnamePort: 443
  ingress:
    enabled: true
    className: "nginx-internal"
    apiDomainName: "dev.eticloud.io"
    domainPrefix: "casa-keycloak"    # → casa-keycloak.dev.eticloud.io
    annotations:
      cert-manager.io/cluster-issuer: letsencrypt

keycloak.hostname must match the ingress hostname exactly — Keycloak uses it to build redirect URIs.

Other common overrides

authService:
  replicaCount: 3

postgresAuth:
  persistence:
    enabled: false   # when using an external PostgreSQL instance

postgresKeycloak:
  persistence:
    enabled: false   # when using an external PostgreSQL instance

sidecar:
  enabled: true      # deploy the ext-auth-service subchart

Full example values file

# values-prod.yaml
authService:
  replicaCount: 3
  database:
    password: "CHANGE_ME"
  idp:
    adminPassword: "CHANGE_ME"
  openai:
    apiBaseUrl: "https://your-litellm-or-openai-endpoint"     # pipeline endpoint
    jwtToken: "CHANGE_ME"                                      # bearer token for pipeline
    llmApiBaseUrl: "https://your-litellm-or-openai-endpoint"  # agent-facing endpoint
    llmApiKey: "CHANGE_ME"                                     # key for agent-facing endpoint
    modelId: "bedrock/global.anthropic.claude-sonnet-4-6"     # model for authorization pipeline
    pipelineModelId: "azure/gpt-4o"                           # model for embedding/tool matching
  ingress:
    enabled: true
    className: "nginx-internal"
    apiDomainName: "dev.eticloud.io"
    domainPrefix: "casa-auth"
    annotations:
      cert-manager.io/cluster-issuer: letsencrypt

uiExplorer:
  ingress:
    enabled: true
    className: "nginx-internal"
    apiDomainName: "dev.eticloud.io"
    domainPrefix: "casa-explorer"
    annotations:
      cert-manager.io/cluster-issuer: letsencrypt

keycloak:
  hostname: "casa-keycloak.dev.eticloud.io"
  hostnamePort: 443
  ingress:
    enabled: true
    className: "nginx-internal"
    apiDomainName: "dev.eticloud.io"
    domainPrefix: "casa-keycloak"
    annotations:
      cert-manager.io/cluster-issuer: letsencrypt

postgresAuth:
  persistence:
    enabled: false

postgresKeycloak:
  persistence:
    enabled: false

sidecar:
  enabled: true

Install with the custom values:

kubectl create namespace casa-dev
kubectl label namespace casa-dev istio-injection=enabled

helm install casa-dev oci://ghcr.io/outshift-open/helm/casa-runtime \
  --version 0.1.17 \
  --namespace casa-dev \
  -f values-prod.yaml

Verify the Installation

Check the auth service health:

kubectl -n casa-dev port-forward svc/casa-dev-auth-service 8000:8000 &
curl http://localhost:8000/health
# Expected: {"status": "healthy"}

Access the UI:

kubectl -n casa-dev port-forward svc/casa-dev-ui-explorer 8080:80 &
# Open http://localhost:8080 in your browser

Apply CRDs

CRDs are included in the Helm chart and installed automatically. Verify they are present:

kubectl get crd | grep casa.io
# multiagentsystems.casa.io
# casapolicies.casa.io

Next Step

Install the Demo MAS →

Or if you're deploying your own MAS, go to Concepts — Multi-Agent Systems.

Install from GHCR (recommended)​

Install from source​

What gets installed​

Production Configuration​

Credentials & LLM backend​

Ingress​

Other common overrides​

Full example values file​

Verify the Installation​

Apply CRDs​

Next Step​