KubeDB MCP Server

A Model Context Protocol server that lets AI agents manage KubeDB databases running in Kubernetes and OpenShift clusters: Postgres, MySQL, MongoDB, Redis, Kafka, Elasticsearch, MariaDB, RabbitMQ, Druid, SingleStore and 25+ other engines.

Written in Go on the official MCP Go SDK. Single static binary, dual transport (stdio for local use, streamable HTTP for in-cluster deployment), discovery driven so it works with any KubeDB release and any subset of installed CRDs.

What agents can do with it

Inventory and observe: list every database across kinds and namespaces with version, phase, replicas, and storage. Pull full manifests, health reports (conditions, pod readiness, warning events), connection endpoints, and the version catalog with upgrade constraints.

Provision and manage lifecycle: create databases with sensible defaults, apply full manifests with server side apply, patch specs, pause and resume, delete with an explicit confirmation gate.

Day-2 operations: every KubeDB OpsRequest type, with dedicated tools for the common paths (restart, version update with catalog validation, horizontal and vertical scaling, volume expansion) and a generic escape hatch for the rest (Reconfigure, ReconfigureTLS, RotateAuth, StorageMigration, Reprovision, database specific types). Plus compute and storage autoscaler configuration.

Data plane resources: schema manager databases (schema.kubedb.com), Kafka ConnectClusters and Connectors, Postgres Publishers and Subscribers, archivers, and Elasticsearch dashboards are all reachable through the generic resource tools.

Tools

Tool	Kind	Description
kubedb_list_kinds	read	Discover installed KubeDB API groups and kinds
kubedb_list_contexts	read	List kubeconfig contexts (multi-cluster)
kubedb_list_databases	read	Database inventory with compact summaries
kubedb_list_resources	read	List any KubeDB family resource
kubedb_get_resource	read	Full manifest of any object, YAML or JSON
kubedb_database_health	read	Phase, conditions, pods, warning events
kubedb_get_connection_info	read	Endpoints, auth secret, TLS state, optional credentials
kubedb_list_versions	read	Catalog versions with deprecation and upgrade constraints
kubedb_list_ops_requests	read	Day-2 operation status
kubedb_create_database	write	Guided provisioning with dryRun
kubedb_apply_manifest	write	Server side apply for any KubeDB manifest
kubedb_patch_resource	write	JSON merge patch or JSON Patch
kubedb_halt_database	write	Pause and resume
kubedb_create_ops_request	write	Generic OpsRequest builder, all types
kubedb_restart_database	write	Safe ordered restart
kubedb_update_version	write	Upgrade with catalog validation
kubedb_scale_database	write	Horizontal and vertical scaling
kubedb_expand_volume	write	Online and offline volume expansion
kubedb_configure_autoscaler	write	Compute and storage autoscaling
kubedb_delete_resource	destructive	Delete with confirm=true gate

Every write tool supports dryRun for server side validation without persisting.

Quick start (local, stdio)

go install kubedb.dev/mcp-server/cmd/kubedb-mcp@latest

Claude Code:

claude mcp add kubedb -- kubedb-mcp

Claude Desktop or any MCP client (uses your current kubeconfig context, switch per call via the context parameter):

{
  "mcpServers": {
    "kubedb": {
      "command": "kubedb-mcp",
      "args": ["--allow-credentials"]
    }
  }
}

Safety model

Flag	Env	Effect
--read-only	KUBEDB_MCP_READ_ONLY	Only the 9 read tools are registered
--disable-destructive	KUBEDB_MCP_DISABLE_DESTRUCTIVE	Removes delete, blocks halt
--allow-credentials	KUBEDB_MCP_ALLOW_CREDENTIALS	Opt-in for credential decoding

Defense in depth beyond the flags: the server refuses to touch any resource outside the *.kubedb.com API groups, deletion requires an explicit confirm=true argument, all tools carry accurate MCP annotations (readOnlyHint, destructiveHint) so clients can apply their own policies, and effective permissions are always bounded by the RBAC of the kubeconfig user or service account.

In-cluster deployment (HTTP)

The container runs the streamable HTTP transport in stateless JSON mode on port 8080, serving MCP at /mcp and health at /healthz.

make image                       # UBI based build via Dockerfile
kubectl create ns kubedb-mcp
kubectl apply -f deploy/openshift/rbac.yaml
kubectl apply -f deploy/openshift/mcpserver.yaml   # MCP lifecycle operator CR

deploy/openshift/mcpserver.yaml targets the MCP lifecycle operator (mcp.x-k8s.io/v1alpha1), the operator behind the OpenShift AI MCP catalog. On clusters without the operator, a plain Deployment plus Service works the same way; the binary needs nothing beyond the service account.

RBAC ships in three tiers: kubedb-mcp-read for observe-only deployments, kubedb-mcp-full for lifecycle and day-2 tools, and kubedb-mcp-credentials (secrets get) only if credential decoding is enabled.

Red Hat OpenShift AI MCP catalog

The MCP catalog in OpenShift AI 3.4+ (AI hub) lists validated MCP servers that admins deploy through the MCP lifecycle operator and consume through the MCP gateway and gen AI studio. This server is built to meet the catalog's technical bar:

1. Streamable HTTP transport, stateless, load balancer friendly (mcp.stateless: true).
2. UBI based image: the Dockerfile builds on ubi9/go-toolset and ships on ubi9-micro, non root (UID 65532), read only root filesystem, restricted SCC compatible, license in /licenses.
3. Required image labels (name, vendor, version, release, summary, description) for Red Hat container certification scanning.
4. Health endpoint at /healthz and MCP at /mcp, matching the operator's defaults.

Onboarding path for the catalog listing:

1. Certify the image through Red Hat Partner Connect container certification (AppsCode already maintains certified KubeDB operator images, so this reuses the existing partner account and pipeline). The certified image publishes to registry.connect.redhat.com.
2. Submit the server for MCP catalog inclusion through the OpenShift AI partner pipeline (partner consent plus technical scanning). Red Hat's validation covers provenance, vulnerability scanning, and transport conformance.
3. Optionally publish server.json to the official MCP registry for discovery outside OpenShift; the file in this repo follows the registry schema.

Development

make build       # static binary in bin/
make vet
make inspector   # interactive testing with the MCP Inspector

License

Apache 2.0. Copyright AppsCode Inc.