diff --git a/nanobot/agent/context.py b/nanobot/agent/context.py
index 3b20af8d..59210d7b 100644
--- a/nanobot/agent/context.py
+++ b/nanobot/agent/context.py
@@ -44,6 +44,10 @@ async def connect_mcp(state: Any, tools: ToolRegistry) -> None:
     await mcp_tools.connect_missing_servers(state, tools)
 
 
+async def close_mcp(state: Any) -> None:
+    await mcp_tools.close_mcp_servers(state)
+
+
 async def handle_runtime_control(state: Any, msg: InboundMessage, tools: ToolRegistry) -> bool:
     return await mcp_tools.handle_runtime_control(state, msg, tools)
 
diff --git a/nanobot/agent/loop.py b/nanobot/agent/loop.py
index 47f29287..21be618e 100644
--- a/nanobot/agent/loop.py
+++ b/nanobot/agent/loop.py
@@ -7,7 +7,7 @@ import dataclasses
 import os
 import time
 from collections.abc import Mapping
-from contextlib import AsyncExitStack, nullcontext, suppress
+from contextlib import nullcontext, suppress
 from dataclasses import dataclass, field
 from enum import Enum, auto
 from functools import partial
@@ -82,6 +82,7 @@ from nanobot.utils.runtime import (
 )
 
 if TYPE_CHECKING:
+    from nanobot.agent.tools.mcp import MCPConnection
     from nanobot.config.schema import (
         ChannelsConfig,
         ProviderConfig,
@@ -360,8 +361,7 @@ class AgentLoop:
         self._unified_session = unified_session
         self._running = False
         self._mcp_servers = mcp_servers or {}
-        self._mcp_stacks: dict[str, AsyncExitStack] = {}
-        self._mcp_retired_stacks: list[tuple[str, AsyncExitStack]] = []
+        self._mcp_stacks: dict[str, MCPConnection] = {}
         self._mcp_connecting = False
         self._active_tasks: dict[str, list[asyncio.Task]] = {}  # session_key -> tasks
         self._background_tasks: list[asyncio.Task] = []
@@ -1178,18 +1178,7 @@ class AgentLoop:
         if self._background_tasks:
             await asyncio.gather(*self._background_tasks, return_exceptions=True)
             self._background_tasks.clear()
-        stacks = [*self._mcp_retired_stacks, *self._mcp_stacks.items()]
-        self._mcp_retired_stacks.clear()
-        for name, stack in stacks:
-            try:
-                await stack.aclose()
-            except asyncio.CancelledError as exc:
-                if not str(exc).startswith("Cancelled via cancel scope"):
-                    raise
-                logger.debug("MCP server '{}' cleanup cancelled by SDK (can be ignored)", name)
-            except (RuntimeError, BaseExceptionGroup):
-                logger.debug("MCP server '{}' cleanup error (can be ignored)", name)
-        self._mcp_stacks.clear()
+        await agent_context.close_mcp(self)
 
     def _schedule_background(self, coro) -> None:
         """Schedule a coroutine as a tracked background task (drained on shutdown)."""
diff --git a/nanobot/agent/tools/mcp.py b/nanobot/agent/tools/mcp.py
index 625127e4..a6d65e69 100644
--- a/nanobot/agent/tools/mcp.py
+++ b/nanobot/agent/tools/mcp.py
@@ -9,7 +9,7 @@ import shutil
 import urllib.parse
 from collections.abc import Awaitable, Callable
 from contextlib import AsyncExitStack, suppress
-from typing import Any, Mapping
+from typing import Any, Mapping, Protocol
 from weakref import WeakKeyDictionary
 
 import httpx
@@ -54,6 +54,26 @@ _RELOAD_LOCKS: WeakKeyDictionary[Any, asyncio.Lock] = WeakKeyDictionary()
 _ReconnectCallback = Callable[[str, str, Tool], Awaitable[Tool | None]]
 
 
+class MCPConnection(Protocol):
+    async def aclose(self) -> None: ...
+
+
+class _OwnedMCPConnection:
+    """Close an MCP transport from the task that originally opened it."""
+
+    def __init__(self, owner: asyncio.Task[None], close_requested: asyncio.Event) -> None:
+        self._owner = owner
+        self._close_requested = close_requested
+
+    async def aclose(self) -> None:
+        self._close_requested.set()
+        try:
+            await asyncio.shield(self._owner)
+        except asyncio.CancelledError:
+            if not self._owner.cancelled():
+                raise
+
+
 def _is_malformed_mcp_progress_notification(message: Any) -> bool:
     payload = _mcp_jsonrpc_payload(message)
     if _payload_value(payload, "method") != "notifications/progress":
@@ -814,19 +834,19 @@ class MCPPromptWrapper(_MCPWrapperBase):
 
 async def connect_mcp_servers(
     mcp_servers: dict, registry: ToolRegistry
-) -> dict[str, AsyncExitStack]:
+) -> dict[str, MCPConnection]:
     """Connect to configured MCP servers and register their tools, resources, prompts.
 
-    Returns a dict mapping server name -> its dedicated AsyncExitStack.
-    Each server gets its own stack to prevent cancel scope conflicts
-    when multiple MCP servers are configured.
+    Returns one connection handle per server.  Each handle keeps the task that
+    entered the MCP SDK contexts alive so reconnect and shutdown can close
+    AnyIO cancel scopes from their owning task.
     """
     from mcp import ClientSession, StdioServerParameters
     from mcp.client.sse import sse_client
     from mcp.client.stdio import stdio_client
     from mcp.client.streamable_http import streamable_http_client
 
-    async def connect_single_server(name: str, cfg) -> tuple[str, AsyncExitStack | None]:
+    async def open_single_server(name: str, cfg) -> tuple[str, AsyncExitStack | None]:
         server_stack = AsyncExitStack()
         await server_stack.__aenter__()
 
@@ -1045,7 +1065,43 @@ async def connect_mcp_servers(
                 await server_stack.aclose()
             return name, None
 
-    server_stacks: dict[str, AsyncExitStack] = {}
+    async def connect_single_server(name: str, cfg) -> tuple[str, MCPConnection | None]:
+        loop = asyncio.get_running_loop()
+        ready: asyncio.Future[bool] = loop.create_future()
+        close_requested = asyncio.Event()
+
+        async def own_connection() -> None:
+            stack: AsyncExitStack | None = None
+            try:
+                _, stack = await open_single_server(name, cfg)
+                if not ready.done():
+                    ready.set_result(stack is not None)
+                if stack is not None:
+                    await close_requested.wait()
+            except BaseException as exc:
+                if not ready.done():
+                    ready.set_exception(exc)
+                raise
+            finally:
+                if stack is not None:
+                    await stack.aclose()
+
+        owner = asyncio.create_task(own_connection(), name=f"mcp:{name}")
+        connection = _OwnedMCPConnection(owner, close_requested)
+        try:
+            connected = await ready
+        except BaseException:
+            close_requested.set()
+            owner.cancel()
+            with suppress(BaseException):
+                await asyncio.shield(owner)
+            raise
+        if not connected:
+            await connection.aclose()
+            return name, None
+        return name, connection
+
+    server_stacks: dict[str, MCPConnection] = {}
 
     for name, cfg in mcp_servers.items():
         try:
@@ -1124,31 +1180,44 @@ def runtime_lines(
 
 async def connect_missing_servers(state: Any, registry: ToolRegistry) -> None:
     """Connect configured MCP servers that are not currently live."""
-    missing_servers = {
-        name: cfg for name, cfg in state._mcp_servers.items() if name not in state._mcp_stacks
-    }
-    if state._mcp_connecting or not missing_servers:
-        return
-    state._mcp_connecting = True
-    try:
-        connected = await connect_mcp_servers(missing_servers, registry)
-        state._mcp_stacks.update(connected)
-        _attach_reconnect_handlers(state, registry, connected)
-        if connected:
-            logger.info("MCP connected servers: {}", sorted(connected))
-        else:
-            logger.warning("No MCP servers connected successfully (will retry next message)")
-    except asyncio.CancelledError:
-        logger.warning("MCP connection cancelled (will retry next message)")
-    except BaseException as e:
-        logger.warning("Failed to connect MCP servers (will retry next message): {}", e)
-    finally:
-        state._mcp_connecting = False
+    async with _reload_lock(state):
+        if getattr(state, "_mcp_closing", False):
+            return
+        missing_servers = {
+            name: cfg for name, cfg in state._mcp_servers.items() if name not in state._mcp_stacks
+        }
+        if state._mcp_connecting or not missing_servers:
+            return
+        state._mcp_connecting = True
+        try:
+            connected = await connect_mcp_servers(missing_servers, registry)
+            if getattr(state, "_mcp_closing", False):
+                for connection in connected.values():
+                    await connection.aclose()
+                return
+            state._mcp_stacks.update(connected)
+            _attach_reconnect_handlers(state, registry, connected)
+            if connected:
+                logger.info("MCP connected servers: {}", sorted(connected))
+            else:
+                logger.warning("No MCP servers connected successfully (will retry next message)")
+        except asyncio.CancelledError:
+            logger.warning("MCP connection cancelled (will retry next message)")
+        except BaseException as e:
+            logger.warning("Failed to connect MCP servers (will retry next message): {}", e)
+        finally:
+            state._mcp_connecting = False
 
 
 async def reload_servers(state: Any, registry: ToolRegistry) -> dict[str, Any]:
     """Reconcile live MCP connections with the current config file."""
     async with _reload_lock(state):
+        if getattr(state, "_mcp_closing", False):
+            return {
+                "ok": False,
+                "message": "MCP connections are shutting down.",
+                "requires_restart": True,
+            }
         try:
             from nanobot.config.loader import load_config, resolve_config_env_vars
 
@@ -1177,7 +1246,7 @@ async def reload_servers(state: Any, registry: ToolRegistry) -> dict[str, Any]:
         tools_removed = 0
         for name in [*removed, *changed]:
             tools_removed += _unregister_server_tools(state, registry, name)
-            _retire_server_stack(state, name)
+            await _close_server(state, name)
 
         state._mcp_servers = next_servers
         retry_missing = sorted(
@@ -1187,9 +1256,17 @@ async def reload_servers(state: Any, registry: ToolRegistry) -> dict[str, Any]:
         )
         to_connect_names = sorted(set(added) | set(changed) | set(retry_missing))
         to_connect = {name: next_servers[name] for name in to_connect_names}
-        connected: dict[str, AsyncExitStack] = {}
+        connected: dict[str, MCPConnection] = {}
         if to_connect:
             connected = await connect_mcp_servers(to_connect, registry)
+            if getattr(state, "_mcp_closing", False):
+                for connection in connected.values():
+                    await connection.aclose()
+                return {
+                    "ok": False,
+                    "message": "MCP connections are shutting down.",
+                    "requires_restart": True,
+                }
             state._mcp_stacks.update(connected)
             _attach_reconnect_handlers(state, registry, connected)
 
@@ -1323,6 +1400,8 @@ async def _refresh_terminated_server(
     stale_tool: Tool,
 ) -> Tool | None:
     async with _reload_lock(state):
+        if getattr(state, "_mcp_closing", False):
+            return None
         cfg = state._mcp_servers.get(server_name)
         if cfg is None:
             logger.warning(
@@ -1341,9 +1420,13 @@ async def _refresh_terminated_server(
 
         logger.warning("MCP server '{}' session terminated; refreshing connection", server_name)
         _unregister_server_tools(state, registry, server_name)
-        _retire_server_stack(state, server_name)
+        await _close_server(state, server_name)
 
         connected = await connect_mcp_servers({server_name: cfg}, registry)
+        if getattr(state, "_mcp_closing", False):
+            for connection in connected.values():
+                await connection.aclose()
+            return None
         state._mcp_stacks.update(connected)
         _attach_reconnect_handlers(state, registry, connected)
         if server_name not in connected:
@@ -1378,17 +1461,24 @@ def _unregister_server_tools(state: Any, registry: ToolRegistry, server_name: st
     return removed
 
 
-def _retire_server_stack(state: Any, server_name: str) -> None:
-    """Remove a stale MCP stack from active use without closing it mid-turn.
-
-    MCP stream transports use AnyIO cancel scopes.  Closing a stack from the
-    reconnecting dispatch task can inject ``CancelledError`` into the task that
-    originally opened it (often ``AgentLoop.run``), which crashes the gateway.
-    Retired stacks are closed later by ``AgentLoop.close_mcp`` during shutdown.
-    """
+async def _close_server(state: Any, server_name: str) -> None:
     stack = state._mcp_stacks.pop(server_name, None)
     if stack is None:
         return
-    retired = getattr(state, "_mcp_retired_stacks", None)
-    if retired is not None:
-        retired.append((server_name, stack))
+    try:
+        await stack.aclose()
+    except (RuntimeError, BaseExceptionGroup):
+        logger.debug("MCP server '{}' cleanup error (can be ignored)", server_name)
+
+
+async def close_mcp_servers(state: Any) -> None:
+    """Close every MCP connection while excluding reconnect and hot reload."""
+    state._mcp_closing = True
+    async with _reload_lock(state):
+        connections = list(state._mcp_stacks.items())
+        state._mcp_stacks.clear()
+        for name, connection in connections:
+            try:
+                await connection.aclose()
+            except (RuntimeError, BaseExceptionGroup):
+                logger.debug("MCP server '{}' cleanup error (can be ignored)", name)
diff --git a/tests/agent/test_mcp_connection.py b/tests/agent/test_mcp_connection.py
index bf8893f3..f5402e84 100644
--- a/tests/agent/test_mcp_connection.py
+++ b/tests/agent/test_mcp_connection.py
@@ -74,15 +74,6 @@ class _FakeMcpTool(Tool):
         return "ok"
 
 
-class _CancelScopeStack:
-    def __init__(self) -> None:
-        self.closed = False
-
-    async def aclose(self) -> None:
-        self.closed = True
-        raise asyncio.CancelledError("Cancelled via cancel scope test")
-
-
 def _make_loop(tmp_path, *, mcp_servers: dict | None = None) -> AgentLoop:
     bus = MessageBus()
     provider = MagicMock()
@@ -125,15 +116,26 @@ async def test_mcp_read_filter_drops_progress_notifications_without_progress_tok
 
 
 @pytest.mark.asyncio
-async def test_close_mcp_swallows_sdk_cancel_scope_cleanup(tmp_path):
-    loop = _make_loop(tmp_path)
-    stack = _CancelScopeStack()
-    loop._mcp_stacks["remote"] = stack  # type: ignore[assignment]
+async def test_owned_mcp_connection_closes_from_its_owner_task():
+    close_requested = asyncio.Event()
+    ready = asyncio.Event()
+    tasks: dict[str, asyncio.Task] = {}
 
-    await loop.close_mcp()
+    async def own_connection() -> None:
+        tasks["open"] = asyncio.current_task()  # type: ignore[assignment]
+        ready.set()
+        await close_requested.wait()
+        tasks["close"] = asyncio.current_task()  # type: ignore[assignment]
 
-    assert stack.closed is True
-    assert loop._mcp_stacks == {}
+    owner = asyncio.create_task(own_connection())
+    connection = mcp_runtime._OwnedMCPConnection(owner, close_requested)
+    await ready.wait()
+
+    await connection.aclose()
+
+    assert tasks["open"] is owner
+    assert tasks["close"] is owner
+    assert tasks["close"] is not asyncio.current_task()
 
 
 @pytest.mark.asyncio
@@ -241,9 +243,6 @@ async def test_reload_mcp_servers_adds_and_removes_tools_without_restart(
     assert removed["removed"] == ["browserbase"]
     assert not loop.tools.has("mcp_browserbase_navigate")
     assert "browserbase" not in loop._mcp_stacks
-
-    assert closed == []
-    await loop.close_mcp()
     assert closed == ["browserbase"]
 
 
@@ -305,9 +304,6 @@ async def test_request_mcp_reload_reaches_runtime_control_without_restart(
     assert result["removed"] == ["browserbase"]
     assert result["requires_restart"] is False
     assert not loop.tools.has("mcp_browserbase_navigate")
-
-    assert closed == []
-    await loop.close_mcp()
     assert closed == ["browserbase"]
 
 
@@ -403,15 +399,12 @@ async def test_mcp_tool_reconnects_after_session_terminated(
 
     assert output == "recovered"
     assert connect_count == 2
-    assert closed == []
+    assert closed == ["remote"]
     assert sessions[0].call_count == 1
     assert sessions[1].call_count == 1
     assert "remote" in loop._mcp_stacks
     assert loop.tools.get("mcp_remote_quote") is not old_tool
 
-    await loop.close_mcp()
-    assert closed == ["remote", "remote"]
-
 
 @pytest.mark.asyncio
 async def test_mcp_reconnect_handler_uses_sanitized_server_prefix(
@@ -521,7 +514,4 @@ async def test_concurrent_mcp_reconnect_reuses_fresh_session(
 
     assert outputs == ["fresh:alpha", "fresh:beta"]
     assert connect_count == 2
-    assert closed == []
-
-    await loop.close_mcp()
-    assert closed == ["remote", "remote"]
+    assert closed == ["remote"]
diff --git a/tests/agent/test_mcp_reconnect_crash.py b/tests/agent/test_mcp_reconnect_crash.py
index ef545283..bea76533 100644
--- a/tests/agent/test_mcp_reconnect_crash.py
+++ b/tests/agent/test_mcp_reconnect_crash.py
@@ -15,7 +15,6 @@ import asyncio
 import multiprocessing
 import socket
 import time
-from contextlib import suppress
 from unittest.mock import MagicMock
 
 import httpx
@@ -70,7 +69,7 @@ async def _wait_for_server(url: str, timeout: float = 10.0) -> bool:
     deadline = time.monotonic() + timeout
     while time.monotonic() < deadline:
         try:
-            async with httpx.AsyncClient(timeout=2.0) as client:
+            async with httpx.AsyncClient(timeout=2.0, trust_env=False) as client:
                 response = await client.get(
                     url,
                     headers={"Accept": "text/event-stream"},
@@ -170,26 +169,30 @@ async def test_mcp_reconnect_after_session_timeout(tmp_path, mcp_server_url):
     )
     loop = _make_loop(tmp_path, mcp_servers={"repro": cfg})
 
-    await loop._connect_mcp()
+    await asyncio.create_task(loop._connect_mcp())
     assert "repro" in loop._mcp_stacks
 
     tool = loop.tools.get("mcp_repro_greet")
     assert isinstance(tool, MCPToolWrapper)
 
-    output = await tool.execute(name="first")
+    output = await asyncio.create_task(tool.execute(name="first"))
     assert "Hello, first" in output
 
     # Wait for the server-side idle timeout to terminate the session.
     await asyncio.sleep(_IDLE_TIMEOUT_SECONDS + 1)
 
-    output = await tool.execute(name="second")
+    output = await asyncio.create_task(tool.execute(name="second"))
     assert "Hello, second" in output
 
-    await loop.close_mcp()
+    await asyncio.create_task(loop.close_mcp())
 
 
 @pytest.mark.asyncio
-async def test_mcp_reconnect_during_shutdown_does_not_crash(tmp_path, mcp_server_url):
+async def test_mcp_reconnect_during_shutdown_does_not_crash(
+    tmp_path,
+    mcp_server_url,
+    monkeypatch: pytest.MonkeyPatch,
+):
     """Simulate the production crash: shutdown while reconnect is in flight."""
     cfg = MCPServerConfig(
         type="streamableHttp",
@@ -199,15 +202,28 @@ async def test_mcp_reconnect_during_shutdown_does_not_crash(tmp_path, mcp_server
     )
     loop = _make_loop(tmp_path, mcp_servers={"repro": cfg})
 
-    await loop._connect_mcp()
+    await asyncio.create_task(loop._connect_mcp())
     tool = loop.tools.get("mcp_repro_greet")
     assert isinstance(tool, MCPToolWrapper)
 
-    await tool.execute(name="first")
+    await asyncio.create_task(tool.execute(name="first"))
     await asyncio.sleep(_IDLE_TIMEOUT_SECONDS + 1)
 
+    reconnect_started = asyncio.Event()
+    finish_reconnect = asyncio.Event()
+    real_connect = mcp_module.connect_mcp_servers
+
+    async def gated_connect(*args, **kwargs):
+        reconnect_started.set()
+        await finish_reconnect.wait()
+        return await real_connect(*args, **kwargs)
+
+    monkeypatch.setattr(mcp_module, "connect_mcp_servers", gated_connect)
     call_task = asyncio.create_task(tool.execute(name="second"))
-    loop.stop()
+    await asyncio.wait_for(reconnect_started.wait(), timeout=5)
+    close_task = asyncio.create_task(loop.close_mcp())
+    await asyncio.sleep(0)
+    finish_reconnect.set()
 
     unhandled: list[BaseException] = []
 
@@ -219,13 +235,11 @@ async def test_mcp_reconnect_during_shutdown_does_not_crash(tmp_path, mcp_server
     asyncio.get_running_loop().set_exception_handler(capture_unhandled)
 
     try:
-        await asyncio.wait_for(asyncio.shield(call_task), timeout=15)
+        await asyncio.wait_for(asyncio.gather(call_task, close_task), timeout=15)
     except asyncio.CancelledError:
         unhandled.append(asyncio.CancelledError("main task cancelled by leaked MCP cancel scope"))
     except Exception as exc:
         unhandled.append(exc)
 
-    with suppress(Exception):
-        await loop.close_mcp()
-
     assert not unhandled, f"Unhandled exception leaked during reconnect/shutdown: {unhandled[0]}"
+    assert loop._mcp_stacks == {}