Installing Katie MCP

Katie MCP is a Model Context Protocol server that can be installed into any compatible MCP Client, such as Claude Desktop. It allows AI Language Models to access accurate, live information from your Kubernetes cluster - provided you have deployed the Katie Agent.

For a curated list of clients, check out the Awesome MCP Clients repository. All usage examples provided have been tested with Claude Desktop.

Prerequisites

A working local Docker environment: the Katie MCP Server is supplied from Docker Hub as a pre-packaged container, in AMD64 and ARM64 architectures.
An MCP Client, capable of loading and using MCPs with an AI Language Model, for example Claude Desktop.

Configuring Katie MCP

Using `stdio` transport (Recommended)

Running Katie MCP with the stdio transport is the recommended approach. Most MCP clients can automatically start and stop the container as needed, communicating with it directly over stdio.

To run Katie MCP in this mode, use the following command:

API_KEY=YOUR_API_KEY docker run \
        --rm \
        -i \
        -e API_KEY \
        intergral/katie-mcp \
        -t stdio

and the environment variable API_KEY must be set in the environment.

For Claude Desktop, the relevant MCP configuration block is:

{
    "mcpServers": {
        "katie-production-docker": {
            "env": {
                "API_KEY": "YOUR_API_KEY"
            },
            "command": "docker",
            "args": [ "run",
                "--rm",
                "-i",
                "-e",
                "API_KEY",
                "intergral/katie-mcp",
                "-t",
                "stdio"
            ]
        }
      }
}

MCP as `SSE` (shared) transport

It is also possible to run the docker separately, perhaps in a shared environment, and have MCP clients connect to it using the Server-Sent Events (SSE) transport.

This method has more network and configuration steps, and is not recommended if the stdio method above works for you.

This method has an additional prerequisite:

The npx command must be available. This is part of the Node Package Manager, commonly installed with the NodeJS environment.

The command to run Katie MCP in this mode is:

API_KEY=YOUR_API_KEY docker run \
        --restart always \
        -i \
        -e API_KEY \
        -p 8000:8000
        intergral/katie-mcp \
        -t sse

This will start the SSE listener on port 8000.

For Claude Desktop, the configuration to communicate with this is:

{
    "mcpServers": {
        "katie-production-sse": {
            "command": "npx",
            "args": [
                "-y",
                "mcp-remote",
                "http://localhost:8000/sse"
            ]
        }
      }
}

Example prompts & workflows

Note

Not all prompts will be available to you if running the Agent with restricted (read-only, or custom) privileges.

Simple queries

Prompt	Description
`What clusters are connected?`	See what clusters are connected via their Katie Agents to FusionReactor Cloud.
`What namespaces are available?`	Self-explanatory.
`List all the pods in the 'otel-demo' namespace, along with their resource requirements and usage.`	Self-explanatory.
`Are any pods in trouble?`	Gets a short summary of any pod problems.
`Get the logs for the restarted pod.`	Retrieves and analyzes the logs for the pod that restarted.
`List all pods in the namespace, together with their resource requests and usages as a percentage of the available node capacity. Order by percentage node memory used.`	Produces a table of pods, showing how demanding each is on the node they’re running on, with a usage summary.
`Show me all Helm deployments`	Gets a list of Helm deployments.

More complex operations

Prompt	Description
`Summarize the cluster, thinking about pressure and resource usage.`	Get a summary of the cluster nodes, and a short summary of any problems.
`Are any deployments under-replicated?`	Ascertain whether all deployments have their required number of pods.
`Scale deployment 'quote' to 2 pods.`	Scale a deployment up or down.
`Scale that deployment back again.`	Revert a previous scaling action.
`Cordon node ip-172-20-1-159.eu-west-1.compute.internal`	Cordon (make unavailable for scheduling new pods) a node.
`Uncordon it again`	Self-explanatory.