Nightwatch

On-call fatigue isn’t a staffing problem — it’s a signal-to-noise problem. ninoxAI ingests alert streams from sources like Prometheus, Checkmk, and Zabbix, deduplicates and clusters them into coherent incidents, then launches an autonomous investigation loop using a tool-calling LLM. The agent queries live systems — container state, cluster health, deployment history, metrics dashboards — reasons over what it finds using a ReAct pattern, and surfaces a ranked list of proposed fixes. Nothing executes until a human approves it.

The defining architectural choice is read-only, human-gated remediation. The agent is explicitly prevented from writing to production systems during investigation. This is not a limitation of ambition — the repo describes it as a core design principle. For teams operating under change-management requirements or in environments where an automated action at the wrong moment causes more damage than the original incident, this constraint is the entire point.

The tool fits best when your alert volume is high enough that triage itself is the bottleneck, your infrastructure spans multiple tools that a human has to pivot between manually, and your risk posture requires a human to review before anything changes. It breaks down if you need closed-loop auto-remediation — the agent proposes but never executes, and there is no configuration path around that. Teams that need autonomous fix execution will reach that ceiling on their first real incident and look elsewhere.

On the integration side, the docs describe connectors for Docker, Kubernetes, AWS, Grafana, GitHub, and Git. Deployment is supported via Docker Compose and Kubernetes manifests, and the repo includes a Dockerfile for local inference, making air-gapped operation viable. The codebase is Apache-2.0 licensed with no vendor hosting layer — you run it, you own it.