[Frontend] Add tool filtering support to ToolServer (#29224)

Signed-off-by: Daniel Salib <danielsalib@meta.com>
Co-authored-by: Chauncey <chaunceyjiang@gmail.com>

tests/entrypoints/openai/test_response_api_mcp_tools.py

@@ -4,6 +4,9 @@
 import pytest
 import pytest_asyncio
 from openai import OpenAI
+from openai_harmony import ToolDescription, ToolNamespaceConfig
+
+from vllm.entrypoints.tool_server import MCPToolServer
 
 from ...utils import RemoteOpenAIServer
 
@@ -111,6 +114,48 @@ async def test_mcp_tool_env_flag_enabled(mcp_enabled_client: OpenAI, model_name:
         )
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+async def test_mcp_tool_with_allowed_tools_star(
+    mcp_enabled_client: OpenAI, model_name: str
+):
+    """Test MCP tool with allowed_tools=['*'] to select all available tools.
+
+    This E2E test verifies that the "*" wildcard works end-to-end.
+    See test_serving_responses.py for detailed unit tests of "*" normalization.
+    """
+    response = await mcp_enabled_client.responses.create(
+        model=model_name,
+        input=(
+            "Execute the following code: "
+            "import random; print(random.randint(1, 1000000))"
+        ),
+        instructions=(
+            "You must use the Python tool to execute code. Never simulate execution."
+        ),
+        tools=[
+            {
+                "type": "mcp",
+                "server_label": "code_interpreter",
+                "server_url": "http://localhost:8888",
+                # Using "*" to allow all tools from this MCP server
+                "allowed_tools": ["*"],
+            }
+        ],
+        extra_body={"enable_response_messages": True},
+    )
+    assert response is not None
+    assert response.status == "completed"
+    # Verify tool calls work with allowed_tools=["*"]
+    tool_call_found = False
+    for message in response.output_messages:
+        recipient = message.get("recipient")
+        if recipient and recipient.startswith("python"):
+            tool_call_found = True
+            break
+    assert tool_call_found, "Should have found at least one Python tool call with '*'"
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
 async def test_mcp_tool_env_flag_disabled(mcp_disabled_client: OpenAI, model_name: str):
@@ -159,3 +204,58 @@ async def test_mcp_tool_env_flag_disabled(mcp_disabled_client: OpenAI, model_nam
         assert message.get("author").get("role") != "developer", (
             "No developer messages should be present without a valid tool"
         )
+
+
+def test_get_tool_description():
+    """Test MCPToolServer.get_tool_description filtering logic.
+
+    Note: The wildcard "*" is normalized to None by
+    _extract_allowed_tools_from_mcp_requests before reaching this layer,
+    so we only test None and specific tool filtering here.
+    See test_serving_responses.py for "*" normalization tests.
+    """
+    pytest.importorskip("mcp")
+
+    server = MCPToolServer()
+    tool1 = ToolDescription.new(
+        name="tool1", description="First", parameters={"type": "object"}
+    )
+    tool2 = ToolDescription.new(
+        name="tool2", description="Second", parameters={"type": "object"}
+    )
+    tool3 = ToolDescription.new(
+        name="tool3", description="Third", parameters={"type": "object"}
+    )
+
+    server.harmony_tool_descriptions = {
+        "test_server": ToolNamespaceConfig(
+            name="test_server", description="test", tools=[tool1, tool2, tool3]
+        )
+    }
+
+    # Nonexistent server
+    assert server.get_tool_description("nonexistent") is None
+
+    # None (no filter) - returns all tools
+    result = server.get_tool_description("test_server", allowed_tools=None)
+    assert len(result.tools) == 3
+
+    # Filter to specific tools
+    result = server.get_tool_description(
+        "test_server", allowed_tools=["tool1", "tool3"]
+    )
+    assert len(result.tools) == 2
+    assert result.tools[0].name == "tool1"
+    assert result.tools[1].name == "tool3"
+
+    # Single tool
+    result = server.get_tool_description("test_server", allowed_tools=["tool2"])
+    assert len(result.tools) == 1
+    assert result.tools[0].name == "tool2"
+
+    # No matching tools - returns None
+    result = server.get_tool_description("test_server", allowed_tools=["nonexistent"])
+    assert result is None
+
+    # Empty list - returns None
+    assert server.get_tool_description("test_server", allowed_tools=[]) is None

tests/entrypoints/openai/test_serving_responses.py

@@ -17,6 +17,7 @@ from vllm.entrypoints.context import ConversationContext
 from vllm.entrypoints.openai.protocol import ErrorResponse, ResponsesRequest
 from vllm.entrypoints.openai.serving_responses import (
     OpenAIServingResponses,
+    _extract_allowed_tools_from_mcp_requests,
     extract_tool_types,
 )
 from vllm.entrypoints.tool_server import ToolServer
@@ -254,3 +255,98 @@ class TestValidateGeneratorInput:
         # Should return an ErrorResponse
         assert result is not None
         assert isinstance(result, ErrorResponse)
+
+
+class TestExtractAllowedToolsFromMcpRequests:
+    """Test class for _extract_allowed_tools_from_mcp_requests function"""
+
+    def test_extract_allowed_tools_basic_formats(self):
+        """Test extraction with list format, object format, and None."""
+        from openai.types.responses.tool import McpAllowedToolsMcpToolFilter
+
+        tools = [
+            # List format
+            Mcp(
+                type="mcp",
+                server_label="server1",
+                allowed_tools=["tool1", "tool2"],
+            ),
+            # Object format
+            Mcp(
+                type="mcp",
+                server_label="server2",
+                allowed_tools=McpAllowedToolsMcpToolFilter(
+                    tool_names=["tool3", "tool4"]
+                ),
+            ),
+            # None (no filter)
+            Mcp(
+                type="mcp",
+                server_label="server3",
+                allowed_tools=None,
+            ),
+        ]
+        result = _extract_allowed_tools_from_mcp_requests(tools)
+        assert result == {
+            "server1": ["tool1", "tool2"],
+            "server2": ["tool3", "tool4"],
+            "server3": None,
+        }
+
+    def test_extract_allowed_tools_star_normalization(self):
+        """Test that '*' wildcard is normalized to None (select all tools).
+
+        This is the key test requested by reviewers to explicitly demonstrate
+        that the "*" select-all scenario is handled correctly.
+        """
+        from openai.types.responses.tool import McpAllowedToolsMcpToolFilter
+
+        tools = [
+            # Star in list format
+            Mcp(
+                type="mcp",
+                server_label="server1",
+                allowed_tools=["*"],
+            ),
+            # Star mixed with other tools in list
+            Mcp(
+                type="mcp",
+                server_label="server2",
+                allowed_tools=["tool1", "*"],
+            ),
+            # Star in object format
+            Mcp(
+                type="mcp",
+                server_label="server3",
+                allowed_tools=McpAllowedToolsMcpToolFilter(tool_names=["*"]),
+            ),
+        ]
+        result = _extract_allowed_tools_from_mcp_requests(tools)
+        # All should be normalized to None (allows all tools)
+        assert result == {
+            "server1": None,
+            "server2": None,
+            "server3": None,
+        }
+
+    def test_extract_allowed_tools_filters_non_mcp(self):
+        """Test that non-MCP tools are ignored during extraction."""
+        tools = [
+            Mcp(
+                type="mcp",
+                server_label="server1",
+                allowed_tools=["tool1"],
+            ),
+            LocalShell(type="local_shell"),  # Non-MCP tool should be ignored
+            Mcp(
+                type="mcp",
+                server_label="server2",
+                allowed_tools=["tool2"],
+            ),
+        ]
+        result = _extract_allowed_tools_from_mcp_requests(tools)
+        # Non-MCP tools should be ignored
+        assert result == {
+            "server1": ["tool1"],
+            "server2": ["tool2"],
+        }