RFC: Proxying individual MCP servers from the core MCP server

[wzxxing]

Is this related to an existing feature request or issue?

No response

Summary

As of 2025/05/30 the awslabs/mcp repo contains 34 MCP servers and roughly 242 tools (by counting the occurrence of @mcp.tool), more RFCs of adding MCP servers from other services are in the issue

The user of these MCP servers need to understand the domain they are working on, configure the MCP servers then start their assistant such as Cursor or Amazon Q.

The current core MCP server is really limited, with just a prompt_understanding tool that returns a markdown content with all the possible MCP servers offered in this repository.

This RFC suggests adding a few extra tools to the current core MCP server

Use case

User can just load the core MCP server and start working with any projects. They can even suggest their LLM to load MCP servers

More and more AWS teams will want to contribute to the MCP servers and this tool will be helpful for MCP server discovery.

Use cases

1. I start amazon Q cli with only core MCP server configured.
2. Start asking a question about AWS EKS
3. Amazon Q calls prompt_understanding then understands that awslabs.eks-mcp-server is available
4. Amazon Q call tool require_server("awslabs.eks-mcp-server")
5. The tools/resources/prompts in the new server is available to Amazon Q.
6. Amazon Q calls list_k8s_resources tool from EKS server.

Proposal

Two new tools added to the awslabs.core-mcp-server MCP server.

1. require_server

This tool should use create a FastMCP proxy that launches individual servers from this repository.

@mcp.tool()
async def require_server(server_name: str) -> Literal['Success']:
    """Load the MCP server from awslabs.

    Args:
        server_name: the name of the server to be used. suggested by the tool `prompt_understanding`
    """
    proxy_server = FastMCP.as_proxy(
        MCPConfig(
            mcpServers={
                server_name: StdioMCPServer(
                    command='uvx',
                    args=[server_name],
                )
            }
        )
    )
    mcp.mount(prefix='required', server=proxy_server)
    return 'Success'

2. suggest_servers

Instead of returning a static string from markdown file like prompt_understanding, suggest_servers should suggest valid MCP server names to be required in the require_server tool. There can be many different implementations such as adding tags (AWS service names) to each MCP servers as metadata, or semantic search using a local faiss index.

Out of scope

loading arbitrary MCP servers.

exact implementation of the semantic search.

Potential challenges

1. server conflict when the user already configured a server, then require_server tries to load the same server.
2. different servers have different required environment variables (But luckily, most MCP servers require only AWS related environment variables).
3. unify the command to launch MCP servers, not all servers should be launched with uvx. The core server can have configs about the command to launch each server.
4. After "mounting" a proxy server, the client needs to either call list_tools again, or to handle notification from the server about tool changes.

Dependencies and Integrations

No response

Alternative solutions

Use other MCP servers as python modules in core MCP server. This locks the implementation of other MCP servers in Python.

[theagenticguy]

Very interesting idea. Funny enough I was looking at the FastMCP dynamic binding feature this morning. I'd be open to looking at a POC of this!

Also, we may want to consider different "types" of core servers. Ones targeted for vibe coding, and another for fully autonomous agents (e.g. Strands SDK, Langgraph, etc).

What do you think?

[wzxxing]

@theagenticguy I will share a POC soon, just created my fork and started the work.

Also, we may want to consider different "types" of core servers. Ones targeted for vibe coding, and another for fully autonomous agents (e.g. Strands SDK, Langgraph, etc).

This can be achieved by suggest_servers, if the specialized servers are properly sorted and tagged.

[wzxxing]

I have implemented the feature in my fork:

wzxxing@69cceb5

To test it, you need to declare the capability of the server to support tool/list_changed server notification.

{
  "mcpServers": {
    "awslabs.core-mcp-server": {
      "command": "/Users/wzxxing/work/aws_mcp/.venv/bin/awslabs.core-mcp-server",
      "env": {
        "FASTMCP_LOG_LEVEL": "ERROR"
      },
      "autoApprove": [],
      "disabled": false,
      "capabilities": {
        "tools": {
          "listChanged": true
        }
      }
    }
  }
}

Chatting with Amazon Q (which supports `tool/list_changed` capability)

> use awslabs tool to help me understand the EKS pods and services running in AWS profile "<REDACTED>"
I'll help you understand the EKS pods and services running in your AWS profile "pulse". Let me first use the prompt understanding tool to analyze your request.


🛠️  Using tool: prompt_understanding from mcp server awslabscore_mcp_server
 ⋮
 ● Running prompt_understanding with the param:
 ⋮  {
 ⋮    "arguments": {},
 ⋮    "name": "prompt_understanding"
 ⋮  }
Allow this action? Use 't' to trust (always allow) this tool for the session. [y/n/t]:

> y

 ⋮
 ● Completed in 0.3s
Let me help you understand the EKS pods and services running in your AWS profile "pulse". First, I'll need to suggest and load the appropriate server to assist with this task.


🛠️  Using tool: suggest_servers from mcp server awslabscore_mcp_server
 ⋮
 ● Running suggest_servers with the param:
 ⋮  {
 ⋮    "arguments": {
 ⋮      "keywords": [
 ⋮        "kubernetes",
 ⋮        "eks",
 ⋮        "aws"
 ⋮      ]
 ⋮    },
 ⋮    "name": "suggest_servers"
 ⋮  }
Allow this action? Use 't' to trust (always allow) this tool for the session. [y/n/t]:

> y

 ⋮
 ● Completed in 0.2s



🛠️  Using tool: require_server from mcp server awslabscore_mcp_server
 ⋮
 ● Running require_server with the param:
 ⋮  {
 ⋮    "arguments": {
 ⋮      "server_name": "awslabs.eks-mcp-server"
 ⋮    },
 ⋮    "name": "require_server"
 ⋮  }
Allow this action? Use 't' to trust (always allow) this tool for the session. [y/n/t]:

> y

 ⋮
 ● Completed in 0.12s
I'll help you understand the EKS pods and services running in your AWS profile "pulse". Let me use the AWS CLI to gather information about your EKS clusters and their resources.

First, let's find out what EKS clusters you have in your "pulse" profile:


🛠️  Using tool: use_aws (trusted)
 ⋮
 ● Running aws cli command:

Service name: eks
Operation name: list-clusters
Parameters:
Profile name: pulse
Region: us-east-1
Label: List EKS clusters
 ⋮
 ● Completed in 1.621s
I found one EKS cluster named "<REDACTED>" in your "pulse" profile. Now, let's get more details about this cluster:


🛠️  Using tool: use_aws (trusted)
 ⋮
 ● Running aws cli command:

Service name: eks
Operation name: describe-cluster
Parameters:
- name: "<REDACTED>"
Profile name: pulse
Region: us-east-1
Label: Get cluster details
 ⋮
 ● Completed in 1.454s



🛠️  Using tool: required_list_api_versions from mcp server awslabscore_mcp_server
 ⋮
 ● Running required_list_api_versions with the param:
 ⋮  {
 ⋮    "arguments": {
 ⋮      "cluster_name": "<REDACTED>"
 ⋮    },
 ⋮    "name": "required_list_api_versions"
 ⋮  }
Allow this action? Use 't' to trust (always allow) this tool for the session. [y/n/t]:

> y

......

Maybe my question is a bit too ambiguous, Amazon Q did not eventually give me a satisfying answer, however, you can see it is using the tools from EKS server via Core server. To verify:

> /tools

Tool                                                                 Permission
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔

Built-in:
- execute_bash                                                       * trust read-only commands
- fs_read                                                            * trusted
- fs_write                                                           * not trusted
- report_issue                                                       * trusted
- use_aws                                                            * trust read-only commands

awslabscore_mcp_server (MCP):
- awslabscore_mcp_server___prompt_understanding                      * not trusted
- awslabscore_mcp_server___require_server                            * not trusted
- awslabscore_mcp_server___required_add_inline_policy                * not trusted
- awslabscore_mcp_server___required_apply_yaml                       * not trusted
- awslabscore_mcp_server___required_generate_app_manifest            * not trusted
- awslabscore_mcp_server___required_get_cloudwatch_logs              * not trusted
- awslabscore_mcp_server___required_get_cloudwatch_metrics           * not trusted
- awslabscore_mcp_server___required_get_k8s_events                   * not trusted
- awslabscore_mcp_server___required_get_pod_logs                     * not trusted
- awslabscore_mcp_server___required_get_policies_for_role            * not trusted
- awslabscore_mcp_server___required_list_api_versions                * not trusted
- awslabscore_mcp_server___required_list_k8s_resources               * not trusted
- awslabscore_mcp_server___required_manage_eks_stacks                * not trusted
- awslabscore_mcp_server___required_manage_k8s_resource              * not trusted
- awslabscore_mcp_server___required_search_eks_troubleshoot_guide    * not trusted
- awslabscore_mcp_server___suggest_servers                           * not trusted

Explicitly asking Q to use a particular tool from EKS.

> can you search eks troubleshoot guide for high CPU metrics?
I'll search the EKS troubleshoot guide for information about high CPU metrics in EKS clusters.


🛠️  Using tool: required_search_eks_troubleshoot_guide from mcp server awslabscore_mcp_server
 ⋮
 ● Running required_search_eks_troubleshoot_guide with the param:
 ⋮  {
 ⋮    "arguments": {
 ⋮      "query": "high CPU metrics in EKS cluster"
 ⋮    },
 ⋮    "name": "required_search_eks_troubleshoot_guide"
 ⋮  }
Allow this action? Use 't' to trust (always allow) this tool for the session. [y/n/t]:

> y

 ⋮
 ● Completed in 10.16s



🛠️  Using tool: required_search_eks_troubleshoot_guide from mcp server awslabscore_mcp_server
 ⋮
 ● Running required_search_eks_troubleshoot_guide with the param:
 ⋮  {
 ⋮    "arguments": {
 ⋮      "query": "troubleshooting high CPU usage in EKS pods"
 ⋮    },
 ⋮    "name": "required_search_eks_troubleshoot_guide"
 ⋮  }
Allow this action? Use 't' to trust (always allow) this tool for the session. [y/n/t]:

> y

 ⋮
 ● Completed in 5.845s



🛠️  Using tool: required_search_eks_troubleshoot_guide from mcp server awslabscore_mcp_server
 ⋮
 ● Running required_search_eks_troubleshoot_guide with the param:
 ⋮  {
 ⋮    "arguments": {
 ⋮      "query": "monitoring CPU metrics in EKS with CloudWatch"
 ⋮    },
 ⋮    "name": "required_search_eks_troubleshoot_guide"
 ⋮  }
Allow this action? Use 't' to trust (always allow) this tool for the session. [y/n/t]:

> y

 ⋮
 ● Completed in 8.778s
Based on the EKS troubleshooting guide, here's what you need to know about high CPU metrics in EKS:

## Monitoring CPU Metrics in EKS

1. Using Kubernetes Metrics Server:
   • Deploy the Metrics Server if not already installed:

     kubectl apply -f https://github.com/kubernetes-sigs/metrics-server/releases/latest/download/components.yaml

   • View CPU usage with:

     kubectl top nodes
     kubectl top pods


2. Using CloudWatch:
   • For clusters running Kubernetes 1.28+, metrics are automatically available in CloudWatch under AWS/EKS namespace
   • Use the Observability dashboard in the Amazon EKS console
   • Deploy CloudWatch Container Insights for comprehensive metrics collection
   • Consider using AWS Distro for OpenTelemetry (ADOT) for correlated metrics

## Troubleshooting High CPU Usage

When you observe high CPU usage in your EKS pods, consider these steps:

1. Resource Management:
   • Set appropriate resource requests and limits for all containers
   • Implement vertical pod autoscaling for efficient resource utilization

2. Application Optimization:
   • Optimize application code and container images
   • Implement proper logging and monitoring to identify resource bottlenecks

3. Health Checks:
   • Use Kubernetes liveness and readiness probes to detect and recover from issues

## Signs of CPU Resource Issues

Look for these symptoms:
• Nodes reporting high CPU utilization
• Increased latency and decreased throughput
• Failed liveness or readiness probes
• Horizontal Pod Autoscaler unable to scale effectively
• Frequent pod restarts

Would you like me to help you implement any specific monitoring solution or troubleshoot a particular high CPU issue in your EKS cluster?

[wzxxing]

I think the prompt_understanding and suggest_servers tool have overlaps and can be combined into one tool.

[github-actions]

This issue is now marked as stale because it hasn't seen activity for a while. Add a comment or it will be closed soon. If you wish to exclude this issue from being marked as stale, add the "backlog" label.