TypeError: Cannot read properties of undefined (reading 'toString')

[WillemBerr]

Hi,

I'm trying to use this MCP Server for K8s on MacOS.

I have a GKE Cluster, and I can interact with cluster through Kubectl.

But when I'm asking Claude Desktop to list pods on my cluster, I always have this error:

Error executing code: MCP error -32603: MCP error -32603: Tool execution failed: TypeError: Cannot read properties of undefined (reading 'toString')

My kubeconfig contains only this cluster, have 655 right, so everyone can read it.

I found this error in server log:

2025-03-27T22:46:52.080Z [kubernetes] [info] Message from server: {"jsonrpc":"2.0","id":77,"error":{"code":-32601,"message":"Method not found"}}

Any idea?

Thank you!

[Flux159]

What tools appear in Claude Desktop when you've connected the mcp server? Do you have any other mcp servers connected as well? Also, does "npx --version" in a terminal show a version number?

That error message isn't super helpful, but could you try debugging in CLI w/ mcp-chat (a client that I wrote to make it easier to test): npx mcp-chat --server "npx mcp-server-kubernetes" -p "List pods in the default namespace.".

I get this response (on a Mac w/ kubectl setup with GKE as well):

Running prompt: List pods in the default namespace.
Connected to server "mcp-server-kubernetes" with tools: [
  'cleanup',             'create_deployment',
  'create_namespace',    'create_pod',
  'delete_pod',          'describe_pod',
  'describe_deployment', 'explain_resource',
  'get_events',          'get_logs',
  'install_helm_chart',  'list_api_resources',
  'list_deployments',    'list_namespaces',
  'list_nodes',          'list_pods',
  'list_services',       'uninstall_helm_chart',
  'upgrade_helm_chart',  'port_forward',
  'stop_port_forward'
]
I'll list the pods in the default namespace for you.
[Tool Call] list_pods
{"namespace": "default"}
Result: {
  "content": [
    {
      "type": "text",
      "text": "{\n  \"pods\": [\n    {\n      \"name\": \"custom-deploy-cv61kk6mt4-dcfd755b6-hxwqz\",\n      \"namespace\": \"default\",\n      \"status\": \"Running\",\n      \"createdAt\": \"2025-03-25T23:31:41.000Z\"\n    },\n    {\n      \"name\": \"custom-deploy-cv61kk6mt4-dcfd755b6-qlvrr\",\n      \"namespace\": \"default\",\n      \"status\": \"Running\",\n      \"createdAt\": \"2025-03-25T23:31:41.000Z\"\n    },\n    {\n      \"name\": \"custom-test-5qz3d0fx4qt\",\n      \"namespace\": \"default\",\n      \"status\": \"Running\",\n      \"createdAt\": \"2025-03-25T23:31:36.000Z\"\n    },\n    {\n      \"name\": \"custom-test-dgrjyvv1zev\",\n      \"namespace\": \"default\",\n      \"status\": \"Running\",\n      \"createdAt\": \"2025-03-25T23:34:44.000Z\"\n    },\n    {\n      \"name\": \"custom-test-ouowmij04h\",\n      \"namespace\": \"default\",\n      \"status\": \"Running\",\n      \"createdAt\": \"2025-03-25T23:29:40.000Z\"\n    },\n    {\n      \"name\": \"postgres-operator-649b6bf48-csrtq\",\n      \"namespace\": \"default\",\n      \"status\": \"Running\",\n      \"createdAt\": \"2025-03-12T19:23:05.000Z\"\n    },\n    {\n      \"name\": \"postgres-operator-ui-659b45d69c-wfm9v\",\n      \"namespace\": \"default\",\n      \"status\": \"Running\",\n      \"createdAt\": \"2025-03-12T19:19:38.000Z\"\n    }\n  ]\n}"
    }
  ]
}
Here are the pods running in the default namespace:

...

[WillemBerr]

Hey @Flux159,

Here is the output of npx mcp-chat --server "npx mcp-server-kubernetes" -p "List pods in the default namespace." :

npx mcp-chat --server "npx mcp-server-kubernetes" -p "List pods in the default namespace."
Running prompt: List pods in the default namespace.
Connected to server "mcp-server-kubernetes" with tools: [
  'cleanup',              'create_deployment',
  'create_namespace',     'create_pod',
  'create_cronjob',       'delete_pod',
  'describe_cronjob',     'describe_pod',
  'describe_deployment',  'explain_resource',
  'get_events',           'get_job_logs',
  'get_logs',             'install_helm_chart',
  'list_api_resources',   'list_cronjobs',
  'list_deployments',     'list_jobs',
  'list_namespaces',      'list_nodes',
  'list_pods',            'list_services',
  'uninstall_helm_chart', 'upgrade_helm_chart',
  'port_forward',         'stop_port_forward',
  'scale_deployment'
]

As you can see, I don't have any answer...

In Claude Desktop, It seems everything is ok.

It Works well with Azure and other On-Prem clusters, but not with GKE...

If I try with MCP Inspector, I can have a great answer (attached file)!

output-mcp-inspector.json

Is there any relevant log file to inspect?
Just checked in "mcp-server-kubernetes.log", I have this error for any requests with GKE Cluster.

2025-03-31T11:53:40.681Z [kubernetes] [info] Message from server: {"jsonrpc":"2.0","id":34,"error":{"code":-32601,"message":"Method not found"}}

thank you!

[Flux159]

Okay it's super weird that it works with Azure & other clusters but not with GKE, and that mcp-chat isn't even displaying an error... - I was thinking it could be something wrong with your authentication to GKE via your kube config, but according to that json it works with MCP inspector. Just as a sanity check, kubectl get pods when connected to your GKE context doesn't display any gke plugin error right?

Could you add a console.log here, rebuild and see if it displays anything? We don't have any logs normally on stdout because it breaks the stdio transport, but for debugging it should at least help you debug. The other option is to check what's happening with a console log before that try block here.

[or-ben-dahan]

@Flux159 Im also getting the exact same error, tried with EKS, GKE and Docker desktop local K8s cluster also.
As you can see k get pods return few pods, Claude Desktop K8s MCP configured and running and the MCP error -32603 error of course..

Macos 15.0
Claude Desktop - Version 0.9.1

[Flux159]

@or-ben-dahan When you typed "I changed it to my local Docker desktop context cluster" - did you change the context via kubectl / kubectx or did you change the config file?

Changing the config file won't work due to the kubernetes-client/javascript library not supporting that. Changing the context should work as long as the client library didn't cache that as well. Could you try restarting Claude desktop after making the kube context / config changes just to make sure that it's not cached?

[or-ben-dahan]

@Flux159 Great, seems that Claude desktop require restart between context changes from terminal (kubectl config use-context docker-desktop).

BUT Now it works only for my docker-desktop cluster, I still have an issue with GKE and EKS since they requires SSO (AWS) and gcloud auth although Im already authenticated in terminal and able to list pods using kubectl (and of course restarted terminal before / after changes).

Local cluster works:

GKE error message:

[Flux159]

@or-ben-dahan For GKE & AWS can you verify that you don't have any custom terminal config (venv, env vars) that are not setup when claude starts the kubernetes mcp server? When I was originally testing I had gke-gcloud-auth-plugin only in a devbox environment which caused issues since Claude couldn't start the MCP server in that custom environment.

[rr-paras-patel]

@WillemBerr do you still face this issue otherwise let's close this issue.