Operating a private AI cluster means context-switching between five tools: kubectl for pod status, Grafana for metrics, k9s for live debugging, a log terminal for trustgate audit events, and nothing that understands LLM-specific concerns — how many prompt injections were blocked, which GPU is saturated, which inference backend is slow. Skyward Console collapses this into a single browser-based UI: a plugin for Headlamp (a CNCF Kubernetes SIG project) that surfaces cluster health, Skyward Gate middleware metrics, per-GPU utilization from the NVIDIA GPU Operator, Skyward Mesh node connectivity, and the full structured audit log from Skyward Gate — all without leaving the browser and with access controlled to administrators only.
Five views in a single browser tab replace the multi-tool context-switching required to operate a private AI cluster day-to-day.
Pod status, node conditions, resource utilisation, and recent events — the standard Kubernetes dashboard view provided by Headlamp itself. RBAC-aware: administrators see everything; read-only service accounts see only permitted namespaces. No custom code required for this view.
Real-time metrics from Skyward Gate (trustgate): requests per second, blocked
injection rate, PII entities scrubbed, middleware chain latency breakdown (P50/P95/P99
per stage). Backed by Prometheus metrics exported by trustgate at /metrics. A live
audit event table shows the last N requests with injection confidence scores and PII
entity types, backed by Loki log aggregation.
Per-GPU metrics from the NVIDIA GPU Operator (DCGM exporter → Prometheus): utilisation %, memory used/total, temperature, and active inference model. Grouped by inference backend (Ollama / vLLM pod) so administrators can see which model is consuming which GPU at a glance. No additional instrumentation required — reads existing DCGM metrics.
Headscale node registry (nodes in tailnet, last-seen timestamps, assigned overlay IPs)
and a Hubble-backed service flow map showing live traffic between SkywardAI components.
Blocked flows appear in red. Useful for diagnosing connectivity issues between Chat,
Gate, and inference backends without running hubble observe in a terminal.
The full tamper-evident audit log from Skyward Gate, rendered as a searchable, filterable table. Filter by model, date range, event type (injection blocked, PII scrubbed, output filtered). Each row expands to show the full structured JSON audit record. Backed by Loki; no log data leaves the cluster.