AI Agents

Korrel8r integrates with AI agents via the Model Context Protocol (MCP), providing tools for correlation search, data retrieval, and OpenShift console navigation. See the MCP reference for details of tools.

Openshift Console navigation

Korrel8r connects an AI agent (via MCP) and the OpenShift console (via REST) through a shared session, enabling conversational troubleshooting: the user looks at something in the console, asks the agent a question, and the agent can understand the context and display findings back in the console.

Prerequisites

• An OpenShift cluster with Lightspeed, Cluster Observability Operator, and desired observability stores (Prometheus, Loki, etc.)
• Korrel8r installed via the Cluster Observability Operator — see Getting Started

Enabling agent navigation in the OpenShift console

1. Enable the feature gate — edit the troubleshooting-panel UIPlugin:

apiVersion: observability.openshift.io/v1alpha1
kind: UIPlugin
metadata:
  name: troubleshooting-panel
spec:
  type: TroubleshootingPanel
  troubleshootingPanel:
    enableAgentNavigation: true

2. Create a route to expose the in-cluster korrel8r service:

oc create route reencrypt --service=korrel8r -n openshift-cluster-observability-operator
export KORREL8R_URL=$(oc get routes/korrel8r -n openshift-cluster-observability-operator -o template='https://{{.spec.host}}')

3. Configure the agent to connect via MCP Streamable HTTP, authenticating as the same user as the console session:

export TOKEN=$(oc whoami -t)

{
  "mcpServers": {
    "korrel8r": {
      "type": "streamable-http",
      "url": "<KORREL8R_URL>/mcp",
      "headers": {
        "Authorization": "Bearer <TOKEN>"
      }
    }
  }
}

4. Enable in the console — click the AI icon in the troubleshooting panel toolbar and toggle the switch on. The icon color indicates status: green = connected, red = connection error (details in the AI menu).

Example: conversational troubleshooting

1. User views an unhealthy deployment in the console and asks: “Why is this failing?”
2. Agent calls get_console — sees k8s:Deployment.apps:{"namespace":"myapp","name":"web"}.
3. Agent calls create_neighbors_graph to explore related signals.
4. Agent finds error logs and a firing alert, retrieves and analyzes them.
5. Agent calls show_in_console to display the relevant logs in the console.

Other integrations

See the MCP reference for details of the available tools.

Connecting via MCP HTTP

korrel8r serves Streamable HTTP at /mcp in web mode:

korrel8r web --http :8080

The agent connects to http://<host>:8080/mcp with a bearer token. This mode also serves the REST API at /api/v1alpha1, required for console navigation. See korrel8r web for options.

Connecting via MCP stdio

For agents that launch MCP servers as subprocesses (Claude Code, Claude Desktop, etc.), use korrel8r mcp:

{
  "mcpServers": {
    "korrel8r": {
      "command": "korrel8r",
      "args": ["--config", "/path/to/korrel8r.yaml", "mcp"]
    }
  }
}

Korrel8r uses the current kubectl/oc login credentials.

Example: Claude Code with local korrel8r

oc login <cluster-url>
curl -o korrel8r.yaml https://raw.githubusercontent.com/korrel8r/korrel8r/main/etc/korrel8r/openshift-route.yaml

Add to .claude/settings.json:

{
  "mcpServers": {
    "korrel8r": {
      "command": "korrel8r",
      "args": ["--config", "korrel8r.yaml", "mcp"]
    }
  }
}

Claude Code will discover korrel8r’s tools automatically. Try:

• “What domains does korrel8r know about?”
• “Find all logs related to deployment web in namespace myapp”
• “What is related to this pod? Show me everything within 2 steps.”