OpenShift MCP Server

How It Works

OpenShift Lightspeed acts as an MCP Client that connects to two MCP Servers. When a user asks a question in natural language, the LLM (served by RHOAI vLLM through LiteLLM) reasons about which tool to call, Lightspeed executes the tool against the MCP Server, and the result is returned to the LLM for analysis and response.

Sequence: MCP Tool Calling Flow

Architecture

High-level topology showing how OpenShift Lightspeed connects to the dual MCP servers through LiteLLM, backed by RHOAI vLLM model serving. Each pod uses the OpenShift logo and runs in the same namespace.

Q Quarkus MCP Server

Custom Java server with 19 operational tools for monitoring, deployment, and performance testing. Built with Fabric8 Kubernetes Client.

checkClusterHealth
getPerformanceMetrics
createDeployment
deployDatabase
runKubeBurner
+ 14 more tools

K Kubernetes MCP Server

Official Go-based server from openshift/openshift-mcp-server. Generic CRUD, pod management, Helm operations.

resources_list / get / create / delete
pods_exec / pods_log
helm_install / helm_list
events_list / namespaces_list
nodes_top / pods_top
+ 10 more tools

+ Supporting Services

Included in the Helm chart for a complete AI-assisted operations stack.

LiteLLM Proxy (OpenAI-compatible gateway)
PostgreSQL (LiteLLM backend DB)
MCP Inspector (testing & debug UI)
RHOAI vLLM (model serving)

OpenShift AI Integration

The MCP Server integrates with Red Hat OpenShift AI (RHOAI) for model serving. vLLM serves Qwen3 8B or Granite models with native tool calling support via the hermes parser. The LiteLLM proxy provides an OpenAI-compatible API gateway that routes requests to the vLLM InferenceService.

vLLM InferenceService Configuration for Tool Calling

apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: isvc-qwen3-8b
spec:
  predictor:
    model:
      modelFormat:
        name: vLLM
      runtime: vllm-runtime
      args:
        - --enable-auto-tool-choice
        - --tool-call-parser=hermes
        - --chat-template=/app/data/template_hermes_tool.jinja

The --tool-call-parser=hermes flag enables native function calling. The LLM receives tool definitions from the MCP servers and can autonomously decide which tool to invoke based on the user's natural language query.

Deployment Options

Choose the deployment mode that matches your environment and permissions. The Developer Sandbox edition provides a safe way to explore MCP capabilities with limited cluster access.

Developer Sandbox

Sandbox Edition — Qwen3 8B

Perfect for evaluating MCP Server capabilities in the Red Hat Developer Sandbox. Uses namespace-scoped permissions and the Qwen3 8B model via RHOAI vLLM. Only tools compatible with ServiceAccount namespace-level permissions are available.

Namespace-scoped ServiceAccount (no cluster-admin)
Qwen3 8B model via OpenShift AI / vLLM
Deployment tools (createDeployment, deployDatabase)
Namespace monitoring (monitorDeployments, detectResourceIssues)
Performance testing (runStorageBenchmark, runCpuStressTest)
HPA, Services, NetworkPolicies
Kubernetes MCP (pods, resources, helm — namespace only)
Cluster-wide monitoring (checkClusterHealth, checkNodeConditions)
Node-level tools (checkKubeletStatus, checkCrioStatus)
Cluster density testing (runKubeBurner)

helm install openshift-mcp-server openshift-mcp/openshift-mcp-server \

            -n $NAMESPACE \

            -f values-developer-sandbox.yaml

ⓘ Tools are limited by the ServiceAccount Role permissions. Cluster-wide monitoring and node-level operations require cluster-admin and are not available in this mode.

Full Version

Full Cluster — cluster-admin

Complete deployment with all 40+ tools enabled. Requires cluster-admin ClusterRoleBinding for full access to nodes, metrics API, cross-namespace monitoring, and cluster density testing.

ClusterRoleBinding with cluster-admin
All 19 custom Quarkus MCP tools
All 21 Kubernetes MCP tools
Full cluster health monitoring across all namespaces
Node conditions, kubelet & CRI-O status
Performance metrics API access
KubeBurner cluster density testing
OLSConfig integration for OpenShift Lightspeed
LiteLLM proxy with configurable model backend
MCP Inspector testing UI

helm install openshift-mcp-server openshift-mcp/openshift-mcp-server \

            -n openshift-lightspeed \

            --set namespace=openshift-lightspeed \

            --set serviceAccount.create=true

⚠ Requires cluster-admin. This deployment creates a ClusterRoleBinding granting full cluster access to the MCP Server ServiceAccount.

Helm Chart

Version	App Version	Description
`0.1.0-alpha`	`0.1.0-alpha`	Dual MCP server + Inspector + LiteLLM (experimental)

Add the Helm Repository

helm repo add openshift-mcp https://maximilianoPizarro.github.io/openshift-mcp-server
helm repo update

Full Install (cluster-admin required)

helm install openshift-mcp-server openshift-mcp/openshift-mcp-server \
  --namespace openshift-lightspeed \
  --set namespace=openshift-lightspeed \
  --set serviceAccount.create=true

Developer Sandbox Install (Qwen3 8B, namespace-scoped)

helm install openshift-mcp-server openshift-mcp/openshift-mcp-server \
  --namespace $NAMESPACE \
  -f values-developer-sandbox.yaml

⚠ Developer Sandbox: The ServiceAccount uses a namespace-scoped Role instead of ClusterRoleBinding. Only tools compatible with namespace-level permissions are functional.

Configure with an existing RHOAI model

helm install openshift-mcp-server openshift-mcp/openshift-mcp-server \
  --namespace openshift-lightspeed \
  --set namespace=openshift-lightspeed \
  --set litellm.model.name=qwen3 \
  --set litellm.model.modelId=isvc-qwen3-8b \
  --set litellm.model.apiBase=http://isvc-qwen3-8b-predictor.qwen3-model.svc.cluster.local:8080/v1

Custom Tools (Quarkus Server)

Monitoring (9 tools)

Tool	Description	Scope
`checkClusterHealth`	Overall cluster health, node/pod status, critical issues	cluster-admin
`getPerformanceMetrics`	Node and pod CPU/memory metrics	cluster-admin
`detectResourceIssues`	Pods with high CPU/memory or excessive restarts	namespace
`analyzePodDisruptions`	Evictions, OOM kills, restart patterns	namespace
`checkNodeConditions`	Node conditions, taints, allocatable resources	cluster-admin
`monitorDeployments`	Deployment rollout health and replica status	namespace
`checkKubeletStatus`	Kubelet service status and journal logs	cluster-admin
`checkCrioStatus`	CRI-O container runtime status	cluster-admin
`analyzeJournalctlPodErrors`	Journal log analysis with filters	cluster-admin

Deployment (5 tools)

Tool	Description	Scope
`createDeployment`	Create deployments with custom config	namespace
`deployDatabase`	Deploy PostgreSQL/MySQL/MongoDB/Redis (RHEL images)	namespace
`createHpa`	Configure horizontal pod autoscalers	namespace
`createService`	Create ClusterIP/NodePort/LoadBalancer services	namespace
`createNetworkPolicy`	Create network policies	namespace

Performance Testing (5 tools)

Tool	Description	Scope
`runKubeBurner`	Cluster density testing	cluster-admin
`runStorageBenchmark`	Storage I/O benchmarks with FIO	namespace
`runNetworkTest`	Network throughput with iperf3	namespace
`runCpuStressTest`	CPU/memory stress testing	namespace
`runDatabaseBenchmark`	Database benchmarks	namespace

Kubernetes MCP Tools (Official Server)

Toolset	Tools
Config	`configuration_contexts_list`, `targets_list`, `configuration_view`
Core	`resources_list`, `resources_get`, `resources_create_or_update`, `resources_delete`, `resources_scale`, `pods_list`, `pods_get`, `pods_delete`, `pods_top`, `pods_exec`, `pods_log`, `pods_run`, `namespaces_list`, `projects_list`, `events_list`, `nodes_top`, `nodes_log`
Helm	`helm_install`, `helm_list`, `helm_uninstall`

Screenshots

OpenShift Topology View

MCP Server pods running in the openshift-lightspeed namespace with OpenShift logos

Lightspeed Tool Calling

Qwen3 8B model reasoning about which MCP tool to call (checkClusterHealth)

Cluster Health Analysis

Lightspeed analyzing failed pods and high-restart containers using checkClusterHealth tool results

List Pods (pods_list)

Listing pods in openshift-lightspeed namespace using the pods_list_in_namespace tool

Detect Resource Issues

detectResourceIssues tool identifying pods with excessive restarts and high resource consumption

Deploy Redis Database

deployDatabase tool deploying a Redis instance using official RHEL catalog image

Create HPA (Horizontal Pod Autoscaler)

createHpa tool configuring auto-scaling with CPU/memory targets

Helm List Releases

helm_list tool showing all Helm releases across namespaces

Helm Chart in OpenShift Catalog

OpenShift MCP Server Helm Chart available in the OpenShift Developer Catalog via HelmChartRepository

MCP Inspector - 19 Tools Connected

MCP Inspector connected to the OpenShift MCP Server showing all 19 available tools

LiteLLM - Model Management (Qwen3 8B)

LiteLLM dashboard showing Qwen3 8B model configured as the AI backend

LiteLLM - MCP Servers Configuration

LiteLLM MCP Servers page showing both MCP servers connected via HTTP transport

OLSConfig Integration

Register MCP servers with OpenShift Lightspeed via the OLSConfig custom resource. Requires the MCPServer feature gate enabled.

apiVersion: ols.openshift.io/v1alpha1
kind: OLSConfig
metadata:
  name: cluster
spec:
  featureGates:
    - MCPServer
  mcpServers:
    - name: openshift-mcp-server
      timeout: 30
      url: 'http://openshift-mcp-server.openshift-lightspeed.svc.cluster.local:8080/mcp'
    - name: kubernetes-mcp-server
      timeout: 30
      url: 'http://openshift-mcp-server-k8s-mcp.openshift-lightspeed.svc.cluster.local:8085/mcp'

Technology Stack

Component	Technology
Custom MCP Server	Quarkus 3.27.3 + Fabric8 Kubernetes Client
Official MCP Server	Go (openshift/openshift-mcp-server)
MCP Protocol	Streamable HTTP (POST /mcp)
Model Serving	OpenShift AI (RHOAI) + vLLM with hermes tool-call parser
LLM Models	Qwen3 8B, Granite 3 8B Instruct (configurable)
LLM Gateway	LiteLLM Proxy (OpenAI-compatible API)
Container Runtime	UBI9 OpenJDK 21 Runtime
Helm Chart	v2 with 5 deployments, namespace & cluster-admin modes

Helm Chart Components

Component	Image	Port	Description
openshift-mcp-server	`quay.io/maximilianopizarro/openshift-mcp-server`	8080	Custom Quarkus MCP
kubernetes-mcp-server	`quay.io/redhat-user-workloads/.../openshift-mcp-server`	8085	Official K8s MCP
mcp-inspector	`mcpuse/inspector`	8080	MCP testing UI
litellm	`litellm/litellm-non_root`	4000	OpenAI-compatible proxy
litellm-db	`registry.redhat.io/rhel9/postgresql-15`	5432	PostgreSQL backend