Deploy on Kubernetes with Helm

Install AgentArea with the agentarea chart. Choose Kubernetes over Docker Compose when you need more than one replica of a service, rolling updates, or kernel isolation for the code agents run — the mcpManager.runtimeClass setting only exists on this target. The chart bundles PostgreSQL, Valkey, RustFS, Temporal, Kratos, and OpenFGA so a default install works on an empty cluster. The bundled PostgreSQL is a single StatefulSet with no replication and no backup, and the chart’s own values file marks it as not for production. Decide which dependencies to externalize before you install, not after.

Prerequisites

Kubernetes with batch/v1 Jobs and networking.k8s.io/v1 NetworkPolicy; the chart README states 1.20 or later
Helm 3.8 or later
A default StorageClass, if you keep the bundled PostgreSQL or RustFS
Gateway API CRDs and a Gateway, if you keep gateway_api in mcpManager.features.enabled
See requirements for the full matrix

Steps

1. Add the chart repository

helm repo add agentarea https://agentarea.github.io/helm-charts
helm repo update

To install from a clone instead, run helm dependency build charts/agentarea first — the Valkey subchart is a repository dependency and is not vendored.

2. Decide what the chart should bring with it

Set these in a values.yaml you keep under version control. Keep everything bundled for an evaluation:

# values.yaml — evaluation only
global:
  deploymentEnv: development

Externalize state for anything you intend to keep:

# values.yaml — external PostgreSQL, Redis, and S3
postgresql:
  enabled: false
redis:
  enabled: false
rustfs:
  enabled: false

global:
  database:
    host: postgres.internal.example.com
    port: 5432
    name: agentarea
    sslMode: require
  redis:
    existingSecret: agentarea-redis-url    # key `url`, holds the full REDIS_URL
  storage:
    type: s3
    endpoint: https://s3.us-east-1.amazonaws.com
    publicEndpoint: https://s3.us-east-1.amazonaws.com
    bucket: agentarea-documents
    region: us-east-1

Disabling postgresql does not create the auxiliary databases for you. Temporal, Kratos, and OpenFGA each need their own database on the external instance; the chart’s create-database Jobs run against global.database.host using the credentials in global.secrets.postgresql, so that role needs CREATEDB. global.storage.publicEndpoint matters as soon as the object store is not reachable from the browser. The platform signs presigned URLs for direct browser uploads and downloads; if this is empty it falls back to the internal endpoint, and those URLs resolve to a cluster-only address.

3. Set the isolation boundary for untrusted code

mcpManager.runtimeClass governs both MCP server instance pods and agent sandbox pods — one setting, so a sandbox can never end up less confined than an MCP server. It defaults to "", the node’s default runtime, which gives untrusted code the host kernel.

mcpManager:
  runtimeClass: gvisor          # requires runsc + containerd shim on the nodes
  instanceNetworkPolicy:
    enabled: true               # no-op unless your CNI enforces NetworkPolicy

kata-qemu is the alternative and needs KVM, which managed clusters on virtualized nodes rarely expose. Kubernetes refuses to schedule the pod if the named RuntimeClass does not exist in the cluster, which is the correct failure — it is better than silently running on runc.

4. Install

helm install agentarea agentarea/agentarea \
  --namespace agentarea --create-namespace \
  -f values.yaml \
  --wait --timeout 15m

On install the chart generates the secrets that were not pre-created — PostgreSQL, Redis, and RustFS credentials, the application encryption key, and the two sandbox HMAC secrets — and annotates each with helm.sh/resource-policy: keep so they survive helm uninstall. Subsequent upgrades look the existing secrets up and reuse them rather than rotating them. To supply your own instead, create the Secrets named in global.secrets.* before installing; see secrets backends. The install runs Jobs in this order: create the auxiliary databases, run agentarea-api migrate (the db-migration Job, working directory /app/apps/api), then agentarea-api reconcile (the registryReconcile Job, which seeds the LLM provider, LLM model, and MCP server catalogs from S3). The migration Job has an init container that blocks on a TCP connection to the database, so it sits in Init:0/1 until PostgreSQL accepts connections.

5. Expose the platform

Ingress is off by default. Port-forward to check the install:

kubectl port-forward -n agentarea svc/agentarea-frontend 3000:3000
kubectl port-forward -n agentarea svc/agentarea-backend 8000:8000

For a real deployment, enable ingress and set the public URLs. See networking — several settings besides ingress.hosts have to agree, and getting only the Ingress right leaves authentication broken.

Verify

Confirm every Job succeeded. A running Deployment with a failed migration Job is a broken install that looks healthy.

kubectl get jobs -n agentarea

Each Job should show Completions 1/1. The migration Job has ttlSecondsAfterFinished: 300, so it disappears five minutes after completing — absence is not failure if the pods are up.

kubectl get pods -n agentarea
kubectl logs -n agentarea -l app.kubernetes.io/component=migration

The migration log ends in Migrations completed successfully, or Stamped database to head revision if the schema already existed. Check the API answers:

kubectl port-forward -n agentarea svc/agentarea-backend 8000:8000 &
curl -s http://localhost:8000/health

Confirm the catalogs seeded, otherwise the model picker in the UI is empty:

kubectl logs -n agentarea -l app.kubernetes.io/component=registry-reconcile

Troubleshooting

The db-migration Job stays in Init:0/1. Its wait-for-db init container loops on nc -z <host> <port> and prints Waiting for database... forever. The host comes from global.database.host, or from the bundled PostgreSQL service when that is empty. Check that the value resolves and that the port is open from inside the namespace. Pods stay Pending with FailedScheduling naming a RuntimeClass. The cluster has no RuntimeClass matching mcpManager.runtimeClass. Install the runtime and its RuntimeClass, or set it back to "" and accept that untrusted code shares the host kernel. PVCs stay Pending. No default StorageClass. Set postgresql.persistence.storageClass and rustfs.persistence, or disable both components and use managed services. MCP server instances are created but nothing can reach them. With mcpManager.backend: kubernetes and the gateway_api feature enabled, the manager creates HTTPRoutes attached to mcpManager.gateway.name in mcpManager.gateway.namespace (default envoy-gateway in envoy-gateway-system). If Gateway API CRDs or that Gateway are absent, the routes are never programmed. Install Envoy Gateway or point the values at your own Gateway. Browser uploads fail with a DNS or connection error on a presigned URL. global.storage.publicEndpoint is empty and the presigned URL points at the in-cluster object store address. Set it to a URL the browser can resolve, and set global.storage.cors.allowedOrigins to your frontend origin. The API pod crash-loops on SECRET_MANAGER_ENCRYPTION_KEY. The agentarea-app-secrets Secret is missing its encryption-key entry — usually because a hand-created Secret replaced the generated one without that key.

Start here

Concepts

Guides

Self-host

Reference

API Endpoints

Community

Deploy on Kubernetes with Helm

Deploy on Kubernetes with Helm

Prerequisites

Steps

1. Add the chart repository

2. Decide what the chart should bring with it

3. Set the isolation boundary for untrusted code

4. Install

5. Expose the platform

Verify

Troubleshooting

​Deploy on Kubernetes with Helm

​Prerequisites

​Steps

​1. Add the chart repository

​2. Decide what the chart should bring with it

​3. Set the isolation boundary for untrusted code

​4. Install

​5. Expose the platform

​Verify

​Troubleshooting

​Related

Deploy on Kubernetes with Helm

Prerequisites

Steps

1. Add the chart repository

2. Decide what the chart should bring with it

3. Set the isolation boundary for untrusted code

4. Install

5. Expose the platform

Verify

Troubleshooting

Related