diff --git a/tests/hermes_cli/test_mcp_startup.py b/tests/hermes_cli/test_mcp_startup.py
index 45d80bfdd6..76bd1661e3 100644
--- a/tests/hermes_cli/test_mcp_startup.py
+++ b/tests/hermes_cli/test_mcp_startup.py
@@ -240,8 +240,14 @@ def test_discover_mcp_tools_spawns_only_allowed_servers(monkeypatch):
     }
     seen: dict[str, dict] = {}
 
+    sdk_probes = {"n": 0}
+
+    def _fake_ensure_sdk():
+        sdk_probes["n"] += 1
+        return True
+
     monkeypatch.setattr(mcp_tool, "_load_mcp_config", lambda: dict(servers))
-    monkeypatch.setattr(mcp_tool, "_ensure_mcp_sdk", lambda: True)
+    monkeypatch.setattr(mcp_tool, "_ensure_mcp_sdk", _fake_ensure_sdk)
     monkeypatch.setattr(mcp_tool, "_try_acquire_mcp_discovery_lock", lambda: mcp_tool._LOCK_UNAVAILABLE)
     monkeypatch.setattr(mcp_tool, "_release_mcp_discovery_lock", lambda *_a, **_k: None, raising=False)
 
@@ -262,10 +268,13 @@ def test_discover_mcp_tools_spawns_only_allowed_servers(monkeypatch):
     mcp_tool.discover_mcp_tools(allowed_mcp_names=["terminal", "code-mcp"])
     assert set(seen) == {"code-mcp"}
 
-    # `-t terminal` — no MCP server in the filter: skip the whole MCP load.
+    # `-t terminal` — no MCP server in the filter: skip the whole MCP load,
+    # including the ~260ms `mcp` SDK import.
     seen.clear()
+    sdk_probes["n"] = 0
     assert mcp_tool.discover_mcp_tools(allowed_mcp_names=["terminal"]) == []
     assert seen == {}
+    assert sdk_probes["n"] == 0
 
 
 def test_background_discovery_honors_server_filter(monkeypatch, _reset_mcp_server_filter):
diff --git a/tools/mcp_tool.py b/tools/mcp_tool.py
index 52e7926d2d..fa91f4c09f 100644
--- a/tools/mcp_tool.py
+++ b/tools/mcp_tool.py
@@ -8231,12 +8231,6 @@ def discover_mcp_tools(allowed_mcp_names: Optional[List[str]] = None) -> List[st
         logger.debug("No MCP servers configured")
         return []
 
-    # SDK import is deferred to HERE so a config with zero MCP servers (the
-    # default) never pays the ~260ms `mcp` import on CLI startup.
-    if not _ensure_mcp_sdk():
-        logger.debug("MCP SDK not available -- skipping MCP tool discovery")
-        return []
-
     if allowed_mcp_names is not None:
         # Filter by MCP-server-name match. Built-in toolset names that aren't
         # MCP servers will simply not match — that's fine; they don't need
@@ -8256,6 +8250,13 @@ def discover_mcp_tools(allowed_mcp_names: Optional[List[str]] = None) -> List[st
             logger.debug("No MCP servers in --toolsets filter; skipping MCP load entirely")
             return []
 
+    # SDK import is deferred to HERE so a config with zero MCP servers (the
+    # default) — or a -t/--toolsets filter that keeps none — never pays the
+    # ~260ms `mcp` import on CLI startup.
+    if not _ensure_mcp_sdk():
+        logger.debug("MCP SDK not available -- skipping MCP tool discovery")
+        return []
+
     # Cross-process discovery guard (#62771). A lock loser waits for
     # the holder, then performs its own process-local discovery. If locking is
     # unavailable or the bounded wait expires, preserve the previous