MCP server issue

[likku123]

apiVersion: toolhive.stacklok.dev/v1alpha1
kind: MCPServer
metadata:
  name: mkp
  namespace: mkp
spec:
  image: "ghcr.io/stackloklabs/mkp/server"
  transport: streamable-http
  args:
    - --read-write=false
  port: 8080
  permissionProfile:
    type: builtin
    name: network
  podTemplateSpec:
    spec:
      serviceAccountName: mkp-sa
      containers:
      - name: mcp
  resources:
    limits:
      cpu: "1000m"
      memory: "0.500Gi"
    requests:
      cpu: "200m"
      memory: "0.100Gi"

Above is my mcp spec and till Friday it's working fine. From today morning I see the below message in logs.

TTP transport started for container mkp on port 8080
6:31AM INF MCP server mkp started successfully
6:31AM WRN Warning: Failed to create client manager: failed to create group state store: failed to create state directory: mkdir /tmp/.local: read-only file system
6:31AM INF Press Ctrl+C to stop or wait for container to exit
6:31AM INF MCP server not initialized yet, skipping health check for mkp
6:32AM INF MCP server not initialized yet, skipping health check for mkp
6:32AM INF MCP server not initialized yet, skipping health check for mkp

I see the below message in vscode agent

Failed to validate tool mcp_k8s_list_resources: Error: tool parameters array type must have items. Please open a Github issue for the MCP server or extension which provides this tool

I tried deploying other mcp servers and it's working as expected using toolhive.

Any pointers on this on what went wrong?

Thank you !

[likku123]

Adding the VS code msp.json entry as well

{
	"servers": 
	{
		"k8s": {
			"url": "https://mcp-testing.hacochu.com/mcp"
            }
	"inputs": []
}

[JAORMX]

Are you running v0.2.1 ? there was a similar issue in v0.2.0 IIRC

[JAORMX]

You could use ghcr.io/stackloklabs/mkp/server:0.2.1 as an image reference to make sure.

[likku123]

It's still the same even with this image as well

Error: tool parameters array type must have items. Please open a Github issue for the MCP server or extension which provides this tool

[JAORMX]

Can I see how you configured the ingress resource?

[likku123]

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
name: mkp-mcp-ingress
namespace: mkp
annotations:
nginx.ingress.kubernetes.io/rewrite-target: /
acme.cert-manager.io/http01-edit-in-place: "true"
cert-manager.io/cluster-issuer: "letsencrypt-cluster-issuer"
nginx.ingress.kubernetes.io/proxy-read-timeout: "1800"
spec:
ingressClassName: nginx
rules:

• host: mcp-testing.hacochu.com
  http:
  paths:
  • path: /
    pathType: Prefix
    backend:
    service:
    name: mcp-mkp-proxy
    port:
    number: 8080
    tls:
• hosts:
  • mcp-testing.hacochu.com
    secretName: mkp-mcp-certs

[JAORMX]

Could you use the markdown format for code? it would make it easier to read

[likku123]

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: mkp-mcp-ingress
  namespace: mkp
  annotations:
    nginx.ingress.kubernetes.io/rewrite-target: /
    acme.cert-manager.io/http01-edit-in-place: "true"
    cert-manager.io/cluster-issuer: "letsencrypt-cluster-issuer"
    nginx.ingress.kubernetes.io/proxy-read-timeout: "1800"
spec:
  ingressClassName: nginx
  rules:
  - host: mcp-testing.hacochu.com
    http:
      paths:
      - path: /
        pathType: Prefix
        backend:
          service:
            name: mcp-mkp-proxy
            port:
              number: 8080
  tls:
  - hosts:
    - mcp-testing.hacochu.com
    secretName: mkp-mcp-certs

[likku123]

@JAORMX
Just a casual check to know if the issue is something related to MCP server? or do I need to check any thing from my side.

Thank you !

[likku123]

Sorry, for continuous bug on this. Just stuck in between to showcase the demo on the MCP to a larger group . It would be great if basic functionality is reverted back.
Thank you !

[JAORMX]

What version of the operator are you using? We should have the file system warnings fixed nowadays.

Also here is what my mkp test deployment looks like:

apiVersion: v1
kind: ServiceAccount
metadata:
  name: mkp-sa
  namespace: toolhive-system
---
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRole
metadata:
  name: mkp-cluster-role
rules:
- apiGroups: ["*"]
  resources: ["*"]
  verbs: ["*"]
---
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRoleBinding
metadata:
  name: mkp-cluster-role-binding
roleRef:
  apiGroup: rbac.authorization.k8s.io
  kind: ClusterRole
  name: mkp-cluster-role
subjects:
- kind: ServiceAccount
  name: mkp-sa
  namespace: toolhive-system
---
apiVersion: toolhive.stacklok.dev/v1alpha1
kind: MCPServer
metadata:
  name: mkp
  namespace: toolhive-system
spec:
  image: ghcr.io/stackloklabs/mkp/server:0.2.2
  transport: streamable-http
  targetPort: 8080
  port: 8080
  serviceAccount: mkp-sa

It's very lenient, but it illustrates how to set up a cluster role for it.

I think you need a targetPort as well. That might be the issue.

Here's what my ingress looks like:

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: mkp-mcp-ingress
  namespace: toolhive-system
  annotations:
    nginx.ingress.kubernetes.io/backend-protocol: "HTTP"
    # Enable SSE support for streamable HTTP
    nginx.ingress.kubernetes.io/proxy-buffering: "off"
    nginx.ingress.kubernetes.io/proxy-read-timeout: "3600"
    nginx.ingress.kubernetes.io/proxy-send-timeout: "3600"
    # Rewrite /mkp/mcp to /mcp
    nginx.ingress.kubernetes.io/rewrite-target: /mcp
spec:
  ingressClassName: nginx
  rules:
  - host: 192.168.1.204.nip.io
    http:
      paths:
      - path: /mkp/mcp
        pathType: Exact
        backend:
          service:
            name: mcp-mkp-proxy
            port:
              number: 8080

In my case, I'm serving several MCP servers, each on a dedicated path; that's why I needed the rewrite target rule.

[likku123]

@JAORMX I have picked up the latest MKP yaml example from here and using the exact same manifest file. Still I get the same message. I am using the targetport as specified in the example.

Failed to validate tool mcp_k8s_list_resources: Error: tool parameters array type must have items. Please open a Github issue for the MCP server or extension which provides this tool

I am installing toolhive helm chart via flux-system.

Below are the helm chart version:

toolhive-crds: 0.0.21
toolhive: 0.2.11.

What version of operator you are using? I can revert back to that version and check it.

Also, one thing I would like to mention . I am using copilot agent from vscode using Claude sonnet 4 model

If possible can you please share your mcp.json config file to verify with mine.

[likku123]

I think it's definitely issue with list_resources tool.

I have disabled this tool and ran with other tools it's working like a charm.

Is it possible to reverback all the changes applied to list_resources to check it?

@JAORMX using which client you are testing? Claude or GitHub copilot ?

Also, can you please specify the docker image version I can use before you made the changes?

[JAORMX]

My MCP configuration in Roo Code just looks like this:

    "mkp": {
      "url": "http://192.168.1.204.nip.io/mkp/mcp",
      "type": "streamable-http",
      "alwaysAllow": [
        "list_resources"
      ]
    },

Nothing special there.

Let me try it with vscode

[likku123]

@JAORMX Hopefully this post might give some insights on one of the tool working in Cursor and not working in vscode.