Nothing in the tool set writes: there is no call here that creates, restarts, scales or deletes a resource, so what you hand an agent is limited to what it can read. The audit log and legacy catalog calls degrade rather than fail, returning JSON that carries _source unavailable and a note when a Rancher build does not expose them. The cluster argument on the Kubernetes tools is a Rancher cluster ID.
A read-only server over Rancher, reaching its Norman v3 API, the Kubernetes resources of the clusters it manages, Fleet GitOps state and Harvester virtualisation.
- Cluster inventory with health, Kubernetes version, provider and node count, plus a cross-cluster total of clusters and projects.
- Kubernetes access on a named cluster: get or list any apiVersion and kind, describe-style output with recent events, namespace event lists, pod log tails and a capacity summary drawn from Nodes.
- Fleet GitOps state — GitRepos, bundles, registered clusters, and a drift report of BundleDeployments sitting in Modified state.
- Harvester inventory: virtual machines with status, IP, node and CPU or memory, one VM in full spec, Longhorn-backed volumes, host maintenance and disk status, VM images with download progress, VLAN networks, KubeOVN subnets and VPCs, addons and cluster settings.
- Rancher administration reads: users, global role bindings, auth configs, tokens, cloud credentials, settings, feature flags, node drivers, catalogs, cluster repos and cluster registration tokens.
- Norman schema discovery, which reports the types, collection links and actions this particular Rancher exposes.
A reachable Rancher server and an API token for it. The KubeOVN subnet and VPC tools need the kubeovn-operator addon enabled, and token and cloud credential values come back redacted unless show-sensitive-data is turned on.
One command plus a key — npm install -g rancher-mcp-server, then supply credentials