🐛 Bug: LangGraphHttpAgent with MCPAppsMiddleware failing to register tools in LangGraph FastAPI endpoint

[rinkesh2010rpp]

♻️ Reproduction Steps

1. Backend Setup: Run a LangGraph agent using add_langgraph_fastapi_endpoint (Python) on localhost:8123.
2. MCP Setup: Any MCP server which has APPs I am using an example called "budget-allocator" tool.
3. Initialize LangGraphHttpAgent pointing to the Python backend.
4. Attach MCPAppsMiddleware to the agent instance with the HTTP URL of the MCP server.
5. Add this agent to copilotRuntimeNextJSAppRouterEndpoint
6. Point this agent URL to the chatbar
   ...

✅ Expected Behavior

When the tool run the budget allocator tool or any other tool which has MCP apps, it should render the UI.

❌ Actual Behavior

LLM is not detecting the tools added in the adaptor.

Screenshots

No response

𝌚 CopilotKit Version

|── @copilotkit/react-core@1.54.0
├── @copilotkit/react-ui@1.54.0
├── @copilotkit/runtime@1.54.0

📄 Logs (Optional)

[copilotkit-support-bot]

I've analyzed the codebase and the issue you're describing. The problem likely stems from how LangGraphHttpAgent handles tool registration from JS-side middleware when connecting to a remote LangGraph Python backend.

Details

🕵️ Investigation Findings

1. Direct Re-export: In the current version of the @copilotkit/runtime, LangGraphHttpAgent is directly re-exported from the underlying @ag-ui/langgraph library. Unlike the JS-native LangGraphAgent, it doesn't have the same built-in state-merging logic that bridges JS-side actions into the CopilotKit state automatically.
2. Remote Tool Visibility: When using LangGraphHttpAgent, your LLM is typically running on the remote Python server (via add_langgraph_fastapi_endpoint). For the LLM to "see" tools added by the MCPAppsMiddleware on the JS side, those tools must be explicitly forwarded to the remote agent or handled by the runtime as "external actions".
3. Middleware Application: While the CopilotRuntime attempts to apply middleware via agent.use(), if the remote agent's schema doesn't account for these dynamically injected tools, the Python-side LLM will not include them in its tool definitions.

✅ Solution: Use Global mcpApps Configuration

Instead of manually attaching the middleware to the agent instance, use the global mcpApps configuration in your CopilotRuntime. This allows the runtime to manage the tool injection more effectively across different agent types.

Update your route.ts as follows:

import { CopilotRuntime, copilotRuntimeNextJSAppRouterEndpoint } from "@copilotkit/runtime";
import { LangGraphHttpAgent } from "@copilotkit/runtime/langgraph";
import { NextRequest } from "next/server";

const runtime = new CopilotRuntime({
  agents: {
    // Define your agent here without .use()
    my_agent: new LangGraphHttpAgent({
      url: "http://localhost:8123",
    }),
  },
  // Apply MCP Apps globally or scoped to the agent
  mcpApps: {
    servers: [
      {
        type: "http",
        url: "http://your-mcp-server-url", // e.g., http://localhost:3108/mcp
        // Optional: scope to specific agent
        // agentId: "my_agent" 
      },
    ],
  },
});

export const POST = async (req: NextRequest) => {
  const { handleRequest } = copilotRuntimeNextJSAppRouterEndpoint({
    runtime,
    serviceAdapter: yourAdapter, // Optional if provided in agent
    endpoint: req.nextUrl.pathname,
  });
  return handleRequest(req);
};

💡 Alternative: Forward Tools via State

If the global config doesn't solve it, the LLM on the Python side needs to be aware of the "CopilotKit Actions". Ensure your Python LangGraph node is configured to receive and handle the copilotkit state which includes injected tools.

Python Backend:

# Ensure your state includes the 'copilotkit' key
class AgentState(TypedDict):
    messages: Annotated[Sequence[BaseMessage], add_messages]
    copilotkit: Dict[str, Any] # This will contain the MCP tools

🔗 References

• MCP Apps Documentation
• LangGraph Integration Guide

Confidence Assessment: High. The pattern of using global mcpApps config is the established way in recent 1.x versions (like 1.54.0) to ensure tools are correctly propagated through the runtime to remote agents.

---

Was this helpful?

If this solution worked for you, please click on the appropriate option below to help us improve:

✅ This was helpful | ❌ Need more help

[copilotkit-support-bot]

👋 Thanks for your feedback!

I've escalated this issue to our development team. Someone from our team will review your request and provide additional assistance shortly.

If this is blocking production, please book a meeting with us at https://cal.com/nathan-tarbert-copilotkit/15min