Self-hosted agent (in-VPC)
For data centers that cannot let telemetry leave their network, run the Denpex platform on your own hardware. Two deployment models: air-gapped (zero egress, no API key required) and cloud mode (diagnoses report to api.denpex.com from behind your firewall).
Helm chart (Kubernetes)
The Denpex Helm chart ships a pinned container image with the agent baked in. Deploy as a DaemonSet (one pod per GPU node) with Prometheus metrics on port 9836.
Air-gapped (zero egress)
With DENPEX_LOCAL=1, the agent diagnoses failures entirely on the host using a bundled copy of the engine: the regex pattern tier, the Xid to node-action router, cross-rank cascade localization, and the full 360-class encyclopedia matched by IDF-weighted similarity. No logs, signatures, or heartbeats leave the cluster. The LLM fallback is cloud-only and simply unavailable offline; the deterministic engine covers the large majority of real failures locally.
Monitoring stack
In air-gapped mode, fleet health is served as Prometheus metrics. Point your existing Prometheus or VictoriaMetrics at each node's :9836/metrics endpoint. A starter stack with Prometheus, VictoriaMetrics, and Grafana pre-wired is included in the on-prem kit.
Air-gapped crash diagnoses are saved to a local incident log (DENPEX_INCIDENT_DIR) and never transmitted. Fully disconnected clusters mirror the container image into their own registry.
Prerequisites
- Python 3.8+ on each node (stdlib only, no additional packages)
- Kubernetes 1.24+ for the Helm chart, or any Linux host for the standalone agent
- For air-gapped mode: download denpex_local.py and denpex_patterns.json alongside the agent
- Network access to https://denpex.com/agent/ for the initial download (or mirror the files internally)
- NVIDIA GPU with appropriate drivers (the agent reads GPU info via NVML)