Problem
helm/agent-platform-connectivity/templates/mcp-kubernetes/mcpserver.yaml renders the MCPServer mcp-kubernetes whenever components.mcp-kubernetes and components.muster are on: a family-less singleton (x_mcp-kubernetes_<tool>, no management_cluster argument) labelled agent-platform.giantswarm.io/tool-group: infrastructure. Installations do not use it — their mcp-kubernetes is a separate release, registered by agent-platform-mcps as <cluster>-mcp-kubernetes with spec.family {name: kubernetes, instanceArg: management_cluster} and the muster.giantswarm.io/management-cluster label — so a deployment that runs the bundled component (the local lab today; any single-installation deployment) gets a different tool namespace than the fleet, and one that wants the family shape ends up with two registrations of the same server (the singleton plus an mcps entry), i.e. duplicate tools.
Decision behind this (2026-09-12): the infrastructure servers are families only, everywhere. The lab side is giantswarm/agentlab#168.
Proposed solution
Give the bundled registration the two shapes an installation can want, and nothing else:
mcp-kubernetes.mcpServer.enabled (default true): render the CR or leave the registration to agent-platform-mcps.
mcp-kubernetes.mcpServer.family ({name: kubernetes, instanceArg: management_cluster}, off by default for compatibility) together with mcp-kubernetes.mcpServer.managementCluster (the value the management_cluster argument selects; defaults to the installation's name): when set, the CR carries spec.family, the muster.giantswarm.io/management-cluster label and is named <managementCluster>-mcp-kubernetes, exactly as agent-platform-mcps renders a member.
Alternatively the bundled registration is routed through the chart's agent-platform-mcps.mcpServers list (one rendering path for every infrastructure server). Either way the family block and the tool-group label follow agent-platform-mcps' conventions (muster.families.<group>.toolGroup).
Acceptance criteria
Problem
helm/agent-platform-connectivity/templates/mcp-kubernetes/mcpserver.yamlrenders the MCPServermcp-kuberneteswhenevercomponents.mcp-kubernetesandcomponents.musterare on: a family-less singleton (x_mcp-kubernetes_<tool>, nomanagement_clusterargument) labelledagent-platform.giantswarm.io/tool-group: infrastructure. Installations do not use it — their mcp-kubernetes is a separate release, registered by agent-platform-mcps as<cluster>-mcp-kuberneteswithspec.family {name: kubernetes, instanceArg: management_cluster}and themuster.giantswarm.io/management-clusterlabel — so a deployment that runs the bundled component (the local lab today; any single-installation deployment) gets a different tool namespace than the fleet, and one that wants the family shape ends up with two registrations of the same server (the singleton plus an mcps entry), i.e. duplicate tools.Decision behind this (2026-09-12): the infrastructure servers are families only, everywhere. The lab side is giantswarm/agentlab#168.
Proposed solution
Give the bundled registration the two shapes an installation can want, and nothing else:
mcp-kubernetes.mcpServer.enabled(defaulttrue): render the CR or leave the registration to agent-platform-mcps.mcp-kubernetes.mcpServer.family({name: kubernetes, instanceArg: management_cluster}, off by default for compatibility) together withmcp-kubernetes.mcpServer.managementCluster(the value themanagement_clusterargument selects; defaults to the installation's name): when set, the CR carriesspec.family, themuster.giantswarm.io/management-clusterlabel and is named<managementCluster>-mcp-kubernetes, exactly as agent-platform-mcps renders a member.Alternatively the bundled registration is routed through the chart's
agent-platform-mcps.mcpServerslist (one rendering path for every infrastructure server). Either way the family block and the tool-group label follow agent-platform-mcps' conventions (muster.families.<group>.toolGroup).Acceptance criteria
<cluster>-mcp-kuberneteswith the family block does not breakplatform-test-style callers that opt in (The lab registers its own cluster as the infrastructure families (kubernetes, prometheus; capi with #56) the way an installation does — the Infrastructure preset resolves to x_kubernetes_*/x_prometheus_* with management_cluster=agentlab, and the fake fleet retires agentlab#168 tracks the lab side)make verifyare unaffected by the default