fix(mcp): semantic tool filter now transforms tools to Chat Completions format

litellm/proxy/hooks/mcp_semantic_filter/hook.py

@@ -105,7 +105,9 @@ async def _expand_mcp_tools(
             openai_tools,
             _,
         ) = await LiteLLM_Proxy_MCP_Handler._process_mcp_tools_to_openai_format(
-            user_api_key_auth=user_api_key_dict, mcp_tools_with_litellm_proxy=mcp_tools
+            user_api_key_auth=user_api_key_dict,
+            mcp_tools_with_litellm_proxy=mcp_tools,
+            target_format="chat",
         )
 
         # Convert Pydantic models to dicts for compatibility

litellm/responses/mcp/litellm_proxy_mcp_handler.py

@@ -372,6 +372,7 @@ async def _process_mcp_tools_to_openai_format(
         mcp_tools_with_litellm_proxy: List[ToolParam],
         litellm_trace_id: Optional[str] = None,
         request_tags: Optional[list[str]] = None,
+        target_format: Literal["responses", "chat"] = "responses",
     ) -> tuple[List[Any], dict[str, str]]:
         """
         Centralized method to process MCP tools through the complete pipeline.
@@ -395,8 +396,9 @@ async def _process_mcp_tools_to_openai_format(
             request_tags=request_tags,
         )
 
-        openai_tools = LiteLLM_Proxy_MCP_Handler._transform_mcp_tools_to_openai(deduplicated_mcp_tools)
-
+        openai_tools = LiteLLM_Proxy_MCP_Handler._transform_mcp_tools_to_openai(
+            deduplicated_mcp_tools, target_format=target_format
+        )
         return openai_tools, tool_server_map
 
     @staticmethod

tests/mcp_tests/test_semantic_tool_filter_e2e.py

@@ -134,3 +134,49 @@ async def test_e2e_semantic_filter():
         f"✅ E2E test passed: Filtering reduced tools from {len(tools)} to {len(result['tools'])}"
     )
     print(f"   Filtered tools: {[t.name for t in result['tools']]}")
+
+
+@pytest.mark.asyncio
+async def test_expand_mcp_tools_uses_chat_format():
+    """_expand_mcp_tools must request the Chat Completions tool shape
+    ({"type": "function", "function": {...}}), not the flat Responses API
+    shape, since it feeds /v1/chat/completions providers with strict
+    schema validation (e.g. hosted_vllm). Regression test for #32281.
+    """
+    from unittest.mock import AsyncMock, patch
+
+    from litellm.proxy.hooks.mcp_semantic_filter.hook import SemanticToolFilterHook
+
+    fake_tools = [
+        MCPTool(
+            name="matomo-matomo_site_list",
+            description="List sites",
+            inputSchema={"type": "object", "properties": {}},
+        ),
+    ]
+
+    with patch(
+        "litellm.responses.mcp.litellm_proxy_mcp_handler.LiteLLM_Proxy_MCP_Handler._get_mcp_tools_from_manager",
+        new=AsyncMock(return_value=(fake_tools, ["matomo"])),
+    ):
+        hook = SemanticToolFilterHook(Mock())
+        result = await hook._expand_mcp_tools(
+            tools=[
+                {
+                    "type": "mcp",
+                    "server_url": "litellm_proxy/mcp/matomo",
+                    "server_label": "matomo_mcp",
+                    "require_approval": "never",
+                }
+            ],
+            user_api_key_dict=Mock(),
+        )
+
+    assert len(result) == 1
+    tool = result[0]
+    assert tool["type"] == "function"
+    assert "function" in tool, (
+        "Expected the Chat Completions wrapper shape "
+        f"{{'type': 'function', 'function': {{...}}}}, got flat keys: {list(tool.keys())}"
+    )
+    assert tool["function"]["name"] == "matomo-matomo_site_list"

tests/test_litellm/responses/mcp/test_litellm_proxy_mcp_handler.py

@@ -608,6 +608,45 @@ def find_spec(self, fullname, path=None, target=None):
     assert result.returncode == 0, result.stderr
 
 
+@pytest.mark.asyncio
+async def test_process_mcp_tools_to_openai_format_forwards_target_format(monkeypatch):
+    """_process_mcp_tools_to_openai_format must forward target_format to
+    _transform_mcp_tools_to_openai, so callers like the semantic tool
+    filter can request the Chat Completions shape. Regression test for #32281.
+    """
+    captured = {}
+
+    async def fake_process_without_transform(*args, **kwargs):
+        return (["tool"], {"tool": "server"})
+
+    def fake_transform(tools, target_format="responses"):
+        captured["target_format"] = target_format
+        return [{"format": target_format}]
+
+    monkeypatch.setattr(
+        LiteLLM_Proxy_MCP_Handler,
+        "_process_mcp_tools_without_openai_transform",
+        fake_process_without_transform,
+    )
+    monkeypatch.setattr(
+        LiteLLM_Proxy_MCP_Handler,
+        "_transform_mcp_tools_to_openai",
+        staticmethod(fake_transform),
+    )
+
+    chat_tools, tool_server_map = (
+        await LiteLLM_Proxy_MCP_Handler._process_mcp_tools_to_openai_format(
+            user_api_key_auth=None,
+            mcp_tools_with_litellm_proxy=[],
+            target_format="chat",
+        )
+    )
+
+    assert captured["target_format"] == "chat"
+    assert chat_tools == [{"format": "chat"}]
+    assert tool_server_map == {"tool": "server"}
+
+
 def test_extract_tool_call_details_reads_anthropic_tool_use_input():
     """
     Regression test (LIT-4517): an Anthropic tool_use block carries its arguments