Introduce a mechanism to expose existing mcp servers through agents sdk

[jineshshah36]

We recently built out an internal MCP server for our team that access our internal API which works great with cloudflare agents sdk. However, many of the requests we are getting from our team now are to connect to third party apps, many of which have an MCP server. The challenge is that few have http servers that work with Oauth and teaching non-technical members how to setup MCP servers and authenticate and understand the security model is even harder.

I'd love a way for the agent MCP server to be able to connect to an existing MCP server and selectively "proxy" tools. This way we can have a single unified, permissioned set of tools through a single cloudflare MCP server.

[mattzcarey]

I love the idea. Just so I understand it correctly, you are looking for a homebrew solution to MCP Portals. Is there a reason why you can't use these?

I'd love to know a bit more about what client you are using? Is this free access in your org or can you do something like this Extensions API using Cursor.

[jineshshah36]

I tried MCP portals, but I found that many servers don't work with it. For example, I cannot setup the github MCP with it, nor can I enforce custom logic or any other controls. In code, I can, for example, pass each individual dev's PAT to avoid breaking access controls/etc. based on who is logged in via the agents sdk mcp server we have built.

Our primary client is Claude Web. The goal is to have something that requires zero configuration for every member of our company. Most solutions today are still very dev focused. If employees need to edit config files or run docker, it's an immediate no-go. Even having them copy-paste credentials is problematic. We love that we have Cloudlfare Zero Trust enforced OAuth for our custom MCP on the agents sdk. We now want to expose more tools and merge servers so we have a single server with all the tools/custom logic that we'd like.

I think the ideal api surface would be a mechanism to register tools of an MCP server as a part of the agents sdk mcp server with some kind of middleware mechanism to allow us to intercept request/response of the tool calls.

Here's some example code:

export class MyMCP extends McpAgent {
  server = new McpServer();
  github = new MCPClient({ ...githubMCPConfig });

  async init() {
    const githubMCPClient = await github.init();
    this.server.proxyTool(githubMCPClient.tools.getRepo, async ({ input }) => {
      // Return early without calling the tool
      if (input.repoName.includes("forbidden")) {
        return {
          output: { content: [{ text: "You do not have permission" }] },
        };
      }

      // Modify the input before calling the tool
      if (input.repoName.includes("foo")) {
        return { input: { ...input, repoName: "bar" } };
      }

      // Call the tool and return the output
      return { input };
    });
  }
}

[jineshshah36]

Wanted to add that I jankily have this working with posthog's agent toolkit where I am exposing a single tool via our mcp server

I also just saw this tool on Twitter that may be a way to do this:

https://x.com/steipete/status/1986036444666675302?s=12

[kennyj42]

Hey there! I'm the product manager for MCP server portals. Hopefully we can get you to a less janky set up 😁

We have some significant OAuth improvements coming next week that should make it possible to use the Github MCPs (and other problem child MCPs with our current implementation). I will nudge this thread when we're in a better spot. I can also reproduce the Github issue and it'll be one of the first things I test after we make the update!

[jineshshah36]

Thanks @kennyj42