AI Agents MCP

AI agents (MCP)

KubeCove can serve read-only tools to AI agents such as Claude Code, Codex and Cursor through the Model Context Protocol (MCP). Agents get the same health, incident, ownership and GitOps evidence the app shows, without kubectl and without any way to change the cluster. MCP is off by default in the desktop app and on the server.

Tools

Tool	Returns
find_problems	Unhealthy Pods, workloads, Jobs, CronJobs, PersistentVolumeClaims and, without a namespace, Nodes, with their signals and latest Warning event, most severe first. Start here.
list_contexts	Kubeconfig contexts and which one is current.
list_namespaces	Namespaces in a context.
list_resource_kinds	Built-in and custom resource kinds.
list_resources	Compact rows for one kind: status, readiness, restarts, age, health and owner. health=unhealthy keeps only problems.
describe_resource	One resource's labels, owners, spec, status and recent events.
get_resource_yaml	One resource as YAML without managedFields.
get_pod_logs	The last lines of one container's log; previous=true reads the crashed instance.
get_resource_metrics	CPU and memory usage from the metrics API, highest CPU first.
list_argocd_applications	Argo CD Applications with sync status and health.
list_helm_releases	Helm releases with chart, revision and status.

Omit context to use the kubeconfig's current context. list_resource_kinds, list_resources, find_problems and get_resource_metrics return 500 rows per page by default; pass limit and offset to page through the rest. get_resource_metrics pages its pods, workloads and nodes lists separately, so pass list with offset. The other list tools return every row and take no paging arguments. Logs default to the last 200 lines and allow at most 1,000.

Secret values are always redacted. Anyone holding the MCP token can still read what your Kubernetes identity can read, including ConfigMaps and Pod logs, so treat the token like a kubeconfig. Cluster content such as logs, events and annotations can be written by anyone with write access to the cluster; agents are told to treat it as data, not instructions.

Desktop app

1. Open Settings > AI agents and turn on MCP server. KubeCove listens on http://127.0.0.1:47615/mcp while the app is open.

2. Select Copy Claude Code command and run it in a terminal. It registers the endpoint with its bearer token:

   claude mcp add --transport http kubecove http://127.0.0.1:47615/mcp --header "Authorization: Bearer <token>"

   For other agents, select Copy token and configure an HTTP MCP server with the header Authorization: Bearer <token>.

The listener accepts only loopback connections and refuses requests from web pages. Agents read through the app's own caches, so repeated questions about a scope you have open are fast. Change MCP port if 47615 is in use; an error appears in Settings when the port cannot be opened. Rotate token replaces the token immediately, so update every agent you configured.

Server

Start kubecove-server with --mcp to serve the tools at /mcp on its existing listener. The token is in <data-dir>/mcp-token. Startup prints the endpoint and that path, never the token. With --quiet, which the systemd unit and the container images use, it prints nothing. Read the token from that file on the server host, then register the endpoint with your agent, for example:

claude mcp add --transport http kubecove https://kubecove.example.com/mcp --header "Authorization: Bearer <token>"

The MCP token opens only /mcp, and the browser sign-in token, cookies and OIDC sessions do not open /mcp. TLS, --allow-remote and the Host checks apply unchanged. With Cloudflare Access, agents also need an Access service token. --mcp cannot be combined with --oidc-kubernetes-credentials, because agent requests carry no user's identity to forward. To rotate the token, stop the server, delete mcp-token and start it again.

Tool calls are logged as command_completed events with command mcp.<tool>, and requests without the token as request_refused with reason mcp_token. Arguments, results and tokens are never logged.