Kubernetes diagnostics,
grounded in evidence.
Ask operational queries about your cluster in plain English. K8Sage inspects live pod state, log tails, events, and node metrics through sandboxed read-only RBAC.
Engineered for operational safety.
Read-only by design
Every check runs under a dedicated ServiceAccount with get, list, and watch permissions only. Zero risk of cluster mutation.
Evidence tool calling
Five specialized tools inspect pod statuses, event logs, log tails, node capacity, and workload counts in real time.
Token-by-token streaming
Tool execution events and model deltas stream over a single SSE connection. No static loading spinners or delays.
Runs on any cluster
Declarative Kustomize overlays drive local k3d development as well as multi-node cloud clusters with zero manifest drift.
Three-stage diagnostic pipeline.
Ask Natural SRE Queries
Submit operational questions in plain English. “Why is my pod in CrashLoopBackOff?” or “Are any nodes experiencing memory pressure?” No need to manually correlate multiple dashboards.
Gather Real-time Evidence
The agent decides which read-only diagnostic tool to invoke (`pod_status`, `pod_events`, `pod_logs`, `node_status`, `cluster_summary`). Telemetry is fetched live via the sandboxed Evidence service.
Stream Verifiable Answer
The assistant synthesizes a grounded answer citing precise log lines, restart counters, and event timestamps with attached tool chips. Streamed token-by-token over Server-Sent Events.
Sandboxed read-only permissions.
The Evidence microservice queries the Kubernetes API under a dedicated ServiceAccount bound to a custom ClusterRole with exactly three verbs: get, list, and watch.
No create, update, delete, exec, or patch verbs are assigned anywhere in the manifests.
The LLM service (`sage`) has zero direct cluster access; it must request diagnostics through the sandboxed `evidence` API.
Inspect your Kubernetes workloads with zero guesswork.
Launch the live interactive assistant console or apply the manifests directly to your k3s/k3d or cloud cluster.