Sentinel (Helm) - in progress
Lutflow's Kubernetes agent is in progress, with no supported public installation yet
In progress: no supported public installation
Sentinel will be a read-only agent that runs in your Kubernetes cluster and collects GPU and inference-engine metrics for each deployment, to measure its cost per useful token without reading prompts or responses. It has no supported public installation and no self-service dashboard yet. To try it as an early evaluation participant, write to support@lutflow.dev.
To measure a vLLM server you already run, use the SDK today. It needs no Kubernetes agent.
Technical preview (unsupported)
The chart can be downloaded without credentials from oci://ghcr.io/lutflow/charts/lutflow-sentinel (version 0.1.5, checked on 2026-10-15). It is a technical preview: it does not connect to a Lutflow dashboard, it has no support and it can change without notice. Install it only in a test cluster.
The command below uses the chart's local measurement-only mode, kafka.enabled=false: no Kafka, no Schema Registry, no command or pricing consumer and no workload termination. The agent's own metrics stay on its :9090/metrics endpoint.
helm install lutflow-sentinel oci://ghcr.io/lutflow/charts/lutflow-sentinel \
--version 0.1.5 \
--namespace lutflow --create-namespace \
--set kafka.enabled=falseKeep --set kafka.enabled=false: the chart enables Kafka by default and then expects a Kafka cluster.
With these options, version 0.1.5 creates a DaemonSet, a ServiceAccount, a ConfigMap, a Service, and a ClusterRole that can only get, list and watch pods, nodes and namespaces, bound by a ClusterRoleBinding. The DaemonSet mounts the kubelet's pod-resources directory from each node (hostPath). Installing it therefore requires permission to create cluster-wide RBAC.
The chart's values say the agent only discovers pods labeled lutflow.io/enforced: "true". In this mode the label only selects pods to observe; nothing is enforced.
Use a dedicated test namespace that was empty before this preview. Do not reuse a namespace containing other workloads. The last cleanup command deletes everything in that namespace, not just this Helm release.
Check and remove it:
kubectl -n lutflow get daemonset
kubectl -n lutflow logs -l app.kubernetes.io/name=lutflow-sentinel --tail=20
helm uninstall lutflow-sentinel -n lutflow
kubectl delete namespace lutflowNot available
- Exporting telemetry to Kafka or to a Lutflow dashboard.
- Terminating workloads when a budget is exceeded.
Neither is part of this preview or of the current product.