fix: mcp-proxy process leak

src/backend/base/langflow/api/v2/mcp.py

@@ -141,6 +141,8 @@ async def get_servers(
     """Get the list of available servers."""
     import asyncio
 
+    from lfx.base.mcp.util import MCPStdioClient, MCPStreamableHttpClient
+
     server_list = await get_server_list(current_user, session, storage_service, settings_service)
 
     if not action_count:
@@ -150,10 +152,15 @@ async def get_servers(
     # Check all of the tool counts for each server concurrently
     async def check_server(server_name: str) -> dict:
         server_info: dict[str, str | int | None] = {"name": server_name, "mode": None, "toolsCount": None}
+        # Create clients that we control so we can clean them up after
+        mcp_stdio_client = MCPStdioClient()
+        mcp_streamable_http_client = MCPStreamableHttpClient()
         try:
             mode, tool_list, _ = await update_tools(
                 server_name=server_name,
                 server_config=server_list["mcpServers"][server_name],
+                mcp_stdio_client=mcp_stdio_client,
+                mcp_streamable_http_client=mcp_streamable_http_client,
             )
             server_info["mode"] = mode.lower()
             server_info["toolsCount"] = len(tool_list)
@@ -199,6 +206,11 @@ async def check_server(server_name: str) -> dict:
             else:
                 await logger.aexception(f"Error checking server {server_name}: {e}")
                 server_info["error"] = f"Error loading server: {e}"
+        finally:
+            # Always disconnect clients to prevent mcp-proxy process leaks
+            # These clients spawn subprocesses that need to be explicitly terminated
+            await mcp_stdio_client.disconnect()
+            await mcp_streamable_http_client.disconnect()
         return server_info
 
     # Run all server checks concurrently