Fix MCP rate limit quota handling

litellm/proxy/auth/auth_utils.py

@@ -949,12 +949,12 @@ def get_key_mcp_rpm_limit(
     """
     if user_api_key_dict.metadata:
         result = user_api_key_dict.metadata.get("mcp_rpm_limit")
-        if result:
+        if result is not None:
             return result
 
     if user_api_key_dict.team_metadata:
         team_limit = user_api_key_dict.team_metadata.get("mcp_rpm_limit")
-        if team_limit:
+        if team_limit is not None:
             return team_limit
 
     return None

litellm/proxy/hooks/parallel_request_limiter_v3.py

@@ -1727,10 +1727,9 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger):
             descriptors=descriptors,
         )
 
-        # Per-MCP-server rate limits. Only honor mcp_server_name on actual MCP
-        # tool calls; otherwise a normal LLM request could inject it in its body
-        # to consume another server's MCP quota and 429 legitimate tool calls.
-        if call_type == CallTypes.call_mcp_tool.value:
+        # REST MCP calls pass the raw body through this hook before server
+        # resolution; only the later synthetic hook payload may carry this key.
+        if call_type == CallTypes.call_mcp_tool.value and "server_id" not in data:
             mcp_server_name = data.get("mcp_server_name", None)
             self._add_mcp_per_key_rate_limit_descriptor(
                 user_api_key_dict=user_api_key_dict,

tests/test_litellm/proxy/auth/test_auth_utils.py

@@ -14,6 +14,7 @@ from litellm.proxy.auth.auth_utils import (
     abbreviate_api_key,
     check_complete_credentials,
     get_end_user_id_from_request_body,
+    get_key_mcp_rpm_limit,
     get_key_model_rpm_limit,
     get_key_model_tpm_limit,
     get_model_from_request,
@@ -92,6 +93,22 @@ class TestGetKeyModelRpmLimit:
         assert result == {}
 
 
+class TestGetKeyMcpRpmLimit:
+    def test_empty_dict_limits_are_returned(self):
+        key_override = UserAPIKeyAuth(
+            api_key="sk-123",
+            metadata={"mcp_rpm_limit": {}},
+            team_metadata={"mcp_rpm_limit": {"github": 50}},
+        )
+        assert get_key_mcp_rpm_limit(key_override) == {}
+
+        team_empty = UserAPIKeyAuth(
+            api_key="sk-123",
+            team_metadata={"mcp_rpm_limit": {}},
+        )
+        assert get_key_mcp_rpm_limit(team_empty) == {}
+
+
 class TestGetKeyModelTpmLimit:
     """Tests for get_key_model_tpm_limit function."""
 

tests/test_litellm/proxy/hooks/test_parallel_request_limiter_v3.py

@@ -2976,6 +2976,30 @@ def test_mcp_descriptor_skipped_for_non_mcp_request_v3():
     assert _find_descriptor(descriptors, "mcp_per_key") is None
 
 
+def test_mcp_descriptor_skipped_for_raw_rest_body_v3():
+    handler, _ = _make_mcp_handler()
+    user_api_key_dict = UserAPIKeyAuth(
+        api_key=hash_token("sk-mcp-key"),
+        team_id="team-1",
+        metadata={"mcp_rpm_limit": {"github": 5}},
+        team_metadata={"mcp_rpm_limit": {"github": 3}},
+    )
+
+    descriptors = _build_mcp_descriptors(
+        handler,
+        user_api_key_dict,
+        {
+            "server_id": "slack",
+            "name": "demo-tool",
+            "arguments": {},
+            "mcp_server_name": "github",
+        },
+    )
+
+    assert _find_descriptor(descriptors, "mcp_per_key") is None
+    assert _find_descriptor(descriptors, "mcp_per_team") is None
+
+
 def test_mcp_per_team_descriptor_created_from_team_metadata_v3():
     handler, _ = _make_mcp_handler()
     user_api_key_dict = UserAPIKeyAuth(