Your own agent, pointed at the cluster you are looking at.
kdashboard does not ship a model, an API key or a hosted service. It spawns the agent CLI you
already use in a panel at the bottom of the window, and serves it a Kubernetes tool surface
over loopback — the same dispatcher the UI calls, pinned to the same context.
Not installed → the panel says so and links the install instructions
The tool surface
Ten tools that read. Four that change anything, each behind a dialog.
Read · no confirmation
get_current_contextwhich cluster and namespace the window is on
list_resourcesany kind, any namespace, with the list projection
get_resourceone object, full spec and status
get_pod_logswith container, previous, tailLines, sinceSeconds and grep
list_eventsfiltered by object or namespace
top_pods · top_nodeslive usage from metrics-server
list_problemsthe same scan the Problems view runs
get_rightsizingrequests against real usage
query_prometheusonly if you configured one
Mutate · blocks on Approve or Deny
scale_workloadset the replica count on a workload
restart_rolloutthe same annotation kubectl rollout restart writes
delete_podone pod, so the controller recreates it
update_container_resourcesrequests and limits on one container
Agent requests a changeThe AI agent wants to run restart_rollout on kind-kdash-review.Deployment broken/crashloop · rolls 1 pod · no spec changeDenyApprove
Dismissing the dialog, or five minutes without an answer, counts as deny · approval can be turned off per machine in Settings
Quick Actions
Right-click any row: Why is it crashing?, Analyze logs, Diagnose rollout, Optimize resources, Check connectivity — eleven in all, offered per kind. The session starts already pointed at that object.
Investigate with agent
A button on the selected problem, on each alert and on the log control bar, plus Investigate all on the Problems header. It hands the agent the diagnosis it already has instead of making it rediscover it.
Presets
From the Prompts menu of the panel: cluster health check, namespace health check, review warning events, rightsizing review, node capacity & pressure.
Use kdashboard as the MCP server for your other tools
Turn on the external endpoint in Settings and the same tool surface is available on a fixed
localhost port (47831 by default) with a bearer token — with the same in-app approval on
every mutation. Settings has a copy button for each client config.
What we do not do Send your cluster, your logs or your kubeconfig anywhere. Ship a model, proxy a provider, or hold an API key. Let the agent write anything the four tools do not cover. Expose the endpoint off localhost, or without a token.What your own CLI sends to its own provider is between you and that provider