AI SRE Agent for Kubernetes

A self-hosted AI SRE agent that troubleshoots your Kubernetes cluster. It lists pods, reads logs, diagnoses CrashLoopBackOff, and asks a human before restarting anything. ## What it does You ask a question. The agent decides which tools it needs, runs them against your cluster, and answers with what it actually found. A real run from a test cluster: > **You:** Is anything unhealthy in the cluster right now? > > **Agent:** All 9 pods are running and ready with no problems detected. Most recent pods have 0 restarts; the otel-collector has 3 restarts over 289 hours, which is acceptable. Ask a follow-up and it remembers: > **You:** Which pod had the most restarts? > > **Agent:** The tinysystems-otel-collector pod had the most restarts, with 3 restarts over 289 hours of uptime. It answered the second question from memory, without calling a single tool. Conversations are stored in your cluster, keyed by a conversation id you choose, and windowed to the last 50 messages. ## The human gate When the agent decides a workload needs a rolling restart, it can't just do it. The request lands on an approval form. Approve, and the restart runs. Deny, and the agent is told a human said no, and it reports that back instead of pretending it succeeded. Pending approvals survive pod restarts and queue up if several arrive at once. ## The tools it has The agent can call exactly three things: `list_pods` (phases, readiness, restart counts, problem detection), `get_pod_logs` (recent lines from one pod by name), and `restart_workload` (rolling restart of a Deployment, StatefulSet or DaemonSet, behind the approval gate). That is the whole surface. It can't delete anything, can't scale anything, can't read secrets. ## What you need - A Kubernetes cluster with [tiny](https://tinysystems.io/download) installed. Minikube is fine. - An Anthropic API key. You paste it per run; it is never stored in the flow. It runs on Claude Haiku by default, so a typical question costs a fraction of a cent. Pod data and logs go to the Anthropic API as part of the conversation; nothing else leaves your infrastructure, and every run leaves a trace you can replay hop by hop. ## Make it yours It's a flow, not a black box. Open it in the editor and the logic is on the canvas. Swap Slack in for the debug panel, add a PromQL tool, point it at another namespace, or ask Claude Code to do any of that for you.