diff --git a/backend/open_webui/main.py b/backend/open_webui/main.py
index 94aae1804..255c9ace5 100644
--- a/backend/open_webui/main.py
+++ b/backend/open_webui/main.py
@@ -1847,13 +1847,28 @@ async def chat_completion(
                 except Exception:
                     pass
         finally:
+            # Clean up MCP clients.  Shield the entire block from
+            # CancelledError so disconnect() can finish even when the
+            # task is being stopped.  Each client is isolated so one
+            # failure doesn't skip the rest.
             try:
                 if mcp_clients := metadata.get('mcp_clients'):
-                    for client in reversed(mcp_clients.values()):
-                        await client.disconnect()
+
+                    async def _cleanup_mcp():
+                        for client in reversed(list(mcp_clients.values())):
+                            try:
+                                await client.disconnect()
+                            except Exception as e:
+                                log.debug(f'Error disconnecting MCP client: {e}')
+
+                    await asyncio.wait_for(
+                        asyncio.shield(_cleanup_mcp()),
+                        timeout=10.0,
+                    )
+            except asyncio.TimeoutError:
+                log.warning('MCP client cleanup timed out after 10 s')
             except Exception as e:
-                log.debug(f'Error cleaning up: {e}')
-                pass
+                log.debug(f'Error cleaning up MCP clients: {e}')
             # Emit chat:active=false when task completes
             try:
                 if metadata.get('chat_id'):
diff --git a/backend/open_webui/utils/mcp/client.py b/backend/open_webui/utils/mcp/client.py
index 13745aa9f..beb2f1507 100644
--- a/backend/open_webui/utils/mcp/client.py
+++ b/backend/open_webui/utils/mcp/client.py
@@ -1,7 +1,10 @@
 import asyncio
+import logging
 from typing import Optional
 from contextlib import AsyncExitStack
 
+log = logging.getLogger(__name__)
+
 import anyio
 
 from mcp import ClientSession
@@ -136,8 +139,39 @@ class MCPClient:
         return result_dict
 
     async def disconnect(self):
-        # Clean up and close the session
-        await self.exit_stack.aclose()
+        """Clean up and close the session.
+
+        This method is idempotent — calling it multiple times or on a
+        client that was never connected is safe.  It shields the close
+        operation from CancelledError and adds a timeout so a hung MCP
+        server cannot block the event loop indefinitely.
+        """
+        exit_stack = self.exit_stack
+        if exit_stack is None:
+            return
+
+        # Prevent double-close from concurrent callers
+        self.exit_stack = None
+        self.session = None
+
+        try:
+            await asyncio.wait_for(
+                asyncio.shield(exit_stack.aclose()),
+                timeout=5.0,
+            )
+        except asyncio.TimeoutError:
+            log.warning('MCPClient.disconnect() timed out after 5 s')
+        except RuntimeError as exc:
+            # The MCP SDK's streamable_http transport uses anyio task
+            # groups and async generators internally.  When we close
+            # a session that was interrupted mid-flight these can
+            # raise RuntimeError ("aclose(): asynchronous generator is
+            # already running" or "Attempted to exit cancel scope in a
+            # different task").  Swallowing the error here prevents the
+            # orphaned coroutines from spinning at 100 % CPU.
+            log.debug('MCPClient.disconnect() suppressed RuntimeError: %s', exc)
+        except Exception as exc:
+            log.debug('MCPClient.disconnect() error: %s', exc)
 
     async def __aenter__(self):
         await self.exit_stack.__aenter__()