Helm Quickstart

Bare-cluster install of the brain. The reference path lives in local/k8s-bootstrap.sh — run it with --help for the current shape.

Steps

  1. Clone the repo.
  2. Export your env (Postgres URL, embedding provider key) per local/k8s-bootstrap.sh --help.
  3. ./local/k8s-bootstrap.sh --apply -n <namespace> — mints tokens and creates the three Opaque Secrets the charts read.
  4. Provision a vault PVC (or set up NFS via an overlay).
  5. helm install each chart from helm/ into your namespace.
  6. Port-forward svc/agentibrain-brain-api and curl /feed with the bearer in local/.k8s-tokens.

Chart defaults shipped

  • Service URLs are namespace-less and resolve via in-cluster DNS.
  • Optional integration URLs (inference gateway, artifact store) are empty.
  • imagePullSecrets: [] — ghcr.io is public.
  • externalSecret.enabled: false — flip to true to wire in ESO.
  • storageClass: "" — cluster default.
  • extraVolumes: [] — supply NFS or PVC via overlay.

storageClass per cluster

Cluster value
k3s local-path
kind standard
minikube standard
EKS gp2 or gp3
GKE standard or premium-rwo
AKS default or managed-csi

"" lets the cluster’s default StorageClass apply — works on most modern clusters with no override.

Platform overlays

A typical production deploy layers NFS (or another RWX volume), ESO (with your secret store of choice), and a LoadBalancer controller on top of these defaults. See examples/values-overlays/ for sample overlays you can copy into your own deployment repo.

Troubleshooting

  • Pod Pending → check storageClass or that the PVC is bound.
  • embeddings crashloop → Postgres unreachable or pgvector not enabled.
  • /feed returns 401 → token mismatch; re-source local/.k8s-tokens.
  • Empty hot_arcs → no clusters yet; wait for brain-ops to tick.

See also

  • local/README.md — Docker Compose path.
  • docs/DEPLOYMENT.md — multi-source ArgoCD patterns.
  • k8s/argocd/README.md — app-of-apps for full-stack ArgoCD deploy.