harivansh-afk/agentikube: stateful kubernetes pods for ai-agents

Isolated stateful agent sandboxes on Kubernetes

helm install agentikube oci://ghcr.io/harivansh-afk/agentikube \
  -n sandboxes --create-namespace \
  -f my-values.yaml

Create a my-values.yaml with your cluster details:

compute:
  clusterName: my-eks-cluster
storage:
  filesystemId: fs-0123456789abcdef0
sandbox:
  image: my-registry/sandbox:latest

See values.yaml for all options.

The Go CLI handles runtime operations that are inherently imperative:

agentikube create demo --provider openai --api-key key>
agentikube list
agentikube ssh demo
agentikube status
agentikube destroy demo

Build it with go build ./cmd/agentikube or make build.

The Helm chart installs:

• StorageClass (efs-sandbox) backed by your EFS filesystem
• SandboxTemplate defining the pod spec
• NetworkPolicy for ingress/egress rules
• SandboxWarmPool (optional, enabled by default)
• Karpenter NodePool + EC2NodeClass (optional, when compute.type: karpenter)

Each agentikube create then adds a Secret, SandboxClaim, and workspace PVC for that user.

cmd/agentikube/              CLI entrypoint
internal/                    config, manifest rendering, kube helpers
chart/agentikube/            Helm chart
scripts/                     CRD download helper

make build                   # compile CLI
make helm-lint               # lint the chart
make helm-template           # dry-run render
go test ./...                # run tests

• Storage is EFS-only for now
• kubectl must be installed (used by init and ssh)
• Fargate is validated in config but templates only cover Karpenter so far
• k9s is great for browsing sandbox resources