Kubernetes Incident Responder

What they did

kubernetes-mcp-server gives the agent live cluster state: pod status, events, logs, and resource specs. github-mcp-server lets it read the failing deployment's manifests, recent commits, and open issues in the repo. opencode is the terminal agent that reasons over both, edits the manifests, and runs the fix loop, while agentmemory keeps the incident timeline so a follow-up session doesn't restart from zero.

How it went

A diagnosed root cause, a patched manifest committed to a branch, and a pull request with the incident writeup attached.

Worth knowing

Cluster credentials in the agent's environment are powerful; a wrong kubectl-equivalent action can mutate production before you review the diff.

Wiring

Connect kubernetes-mcp-server with a read-scoped kubeconfig first, then github-mcp-server against the repo. Point opencode at the repo working copy and let it call both MCP servers; enable agentmemory so the incident notes persist across sessions.

Try it yourself

The checkout-api pods in namespace prod are crashlooping since the last deploy. Pull the cluster state, find the bad manifest in our repo, fix it, and open a PR with what you found.