[Frontend] Responses API MCP tools for built in tools and to pass through headers

tests/entrypoints/openai/test_response_api_with_harmony.py

@@ -398,7 +398,13 @@ async def test_web_search(client: OpenAI, model_name: str):
 async def test_code_interpreter(client: OpenAI, model_name: str):
     response = await client.responses.create(
         model=model_name,
-        input="Multiply 64548*15151 using builtin python interpreter.",
+        # TODO: Ideally should be able to set max tool calls
+        # to prevent multi-turn, but it is not currently supported
+        # would speed up the test
+        input=("What's the first 4 digits after the decimal point of "
+               "cube root of `19910212 * 20250910`? "
+               "Show only the digits. The python interpreter is not stateful "
+               "and you must print to see the output."),
         tools=[{
             "type": "code_interpreter",
             "container": {
@@ -408,6 +414,32 @@ async def test_code_interpreter(client: OpenAI, model_name: str):
     )
     assert response is not None
     assert response.status == "completed"
+    assert response.usage.output_tokens_details.tool_output_tokens > 0
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.skip(reason="Code interpreter tool is not available in CI yet.")
+async def test_mcp_tool(client: OpenAI, model_name: str):
+    response = await client.responses.create(
+        model=model_name,
+        # TODO: Ideally should be able to set max tool calls
+        # to prevent multi-turn, but it is not currently supported
+        # would speed up the test
+        input=("What's the first 4 digits after the decimal point of "
+               "cube root of `19910212 * 20250910`? "
+               "Show only the digits. The python interpreter is not stateful "
+               "and you must print to see the output."),
+        tools=[{
+            "type": "mcp",
+            "server_label": "code_interpreter",
+            # URL unused for DemoToolServer
+            "server_url": "http://localhost:8888"
+        }],
+    )
+    assert response is not None
+    assert response.status == "completed"
+    assert response.usage.output_tokens_details.tool_output_tokens > 0
 
 
 def get_weather(latitude, longitude):

vllm/entrypoints/context.py

@@ -8,6 +8,7 @@
 from contextlib import AsyncExitStack
 from typing import TYPE_CHECKING, Optional, Union
 
+from openai.types.responses.tool import Mcp
 from openai_harmony import Author, Message, Role, StreamState, TextContent
 
 from vllm.entrypoints.harmony_utils import (
@@ -22,6 +23,21 @@
 logger = logging.getLogger(__name__)
 
 
+# This is currently needed as the tool type doesn't 1:1 match the
+# tool namespace, which is what is used to look up the
+# connection to the tool server
+def _map_tool_name_to_tool_type(tool_name: str) -> str:
+    if tool_name == "browser":
+        return "web_search_preview"
+    elif tool_name == "python":
+        return "code_interpreter"
+    elif tool_name == "container":
+        return "container"
+    else:
+        raise ValueError(
+            f"Built in tool name not defined in mapping: {tool_name}")
+
+
 class TurnTokens:
     """Tracks token counts for a single conversation turn."""
 
@@ -59,8 +75,8 @@ def render_for_completion(self) -> list[int]:
 
     @abstractmethod
     async def init_tool_sessions(self, tool_server: Optional[ToolServer],
-                                 exit_stack: AsyncExitStack,
-                                 request_id: str) -> None:
+                                 exit_stack: AsyncExitStack, request_id: str,
+                                 mcp_tools: dict[str, Mcp]) -> None:
         pass
 
     @abstractmethod
@@ -96,8 +112,8 @@ def render_for_completion(self) -> list[int]:
         raise NotImplementedError("Should not be called.")
 
     async def init_tool_sessions(self, tool_server: Optional[ToolServer],
-                                 exit_stack: AsyncExitStack,
-                                 request_id: str) -> None:
+                                 exit_stack: AsyncExitStack, request_id: str,
+                                 mcp_tools: dict[str, Mcp]) -> None:
         pass
 
     async def cleanup_session(self) -> None:
@@ -315,13 +331,17 @@ async def call_python_tool(self, tool_session: Union["ClientSession",
         ]
 
     async def init_tool_sessions(self, tool_server: Optional[ToolServer],
-                                 exit_stack: AsyncExitStack,
-                                 request_id: str) -> None:
+                                 exit_stack: AsyncExitStack, request_id: str,
+                                 mcp_tools: dict[str, Mcp]):
         if tool_server:
             for tool_name in self.available_tools:
                 if tool_name not in self._tool_sessions:
+                    tool_type = _map_tool_name_to_tool_type(tool_name)
+                    headers = mcp_tools[
+                        tool_type].headers if tool_type in mcp_tools else None
                     tool_session = await exit_stack.enter_async_context(
-                        tool_server.new_session(tool_name, request_id))
+                        tool_server.new_session(tool_name, request_id,
+                                                headers))
                     self._tool_sessions[tool_name] = tool_session
                     exit_stack.push_async_exit(self.cleanup_session)
 

vllm/entrypoints/harmony_utils.py

@@ -126,8 +126,10 @@ def get_developer_message(
         function_tools: list[Union[Tool, ChatCompletionToolsParam]] = []
         for tool in tools:
             if tool.type in ("web_search_preview", "code_interpreter",
-                             "container"):
+                             "container", "mcp"):
                 # These are built-in tools that are added to the system message.
+                # Adding in MCP for now until we support MCP tools executed
+                # server side
                 pass
 
             elif tool.type == "function":

vllm/entrypoints/openai/serving_responses.py

@@ -453,8 +453,12 @@ async def responses_full_generator(
 
         async with AsyncExitStack() as exit_stack:
             try:
+                mcp_tools = {
+                    tool.server_label: tool
+                    for tool in request.tools if tool.type == "mcp"
+                }
                 await context.init_tool_sessions(self.tool_server, exit_stack,
-                                                 request.request_id)
+                                                 request.request_id, mcp_tools)
                 async for _ in result_generator:
                     pass
             except asyncio.CancelledError:
@@ -728,9 +732,13 @@ def _construct_input_messages_with_harmony(
             # New conversation.
             reasoning_effort = (request.reasoning.effort
                                 if request.reasoning else None)
-            # Temporary: OpenAI types doesn't have container tool
-            # so we used MCP to cover that, up for change
-            tool_types = [tool.type for tool in request.tools]
+            # If there is a MCP tool with a matching server_label, then
+            # the tool is enabled. Native types like code_interpreter work
+            # as well
+            tool_types = [
+                tool.server_label if tool.type == "mcp" else tool.type
+                for tool in request.tools
+            ]
             if envs.VLLM_GPT_OSS_USE_CONTAINER_TOOL:
                 tool_types.append("container")
             enable_browser = ("web_search_preview" in tool_types
@@ -1654,8 +1662,12 @@ def _increment_sequence_number_and_return(event: T) -> T:
         async with AsyncExitStack() as exit_stack:
             processer = None
             if self.use_harmony:
+                mcp_tools = {
+                    tool.server_label: tool
+                    for tool in request.tools if tool.type == "mcp"
+                }
                 await context.init_tool_sessions(self.tool_server, exit_stack,
-                                                 request.request_id)
+                                                 request.request_id, mcp_tools)
                 processer = self._process_harmony_streaming_events
             else:
                 processer = self._process_simple_streaming_events

vllm/entrypoints/tool_server.py

@@ -18,7 +18,6 @@
 async def list_server_and_tools(server_url: str):
     from mcp import ClientSession
     from mcp.client.sse import sse_client
-
     async with sse_client(url=server_url) as streams, ClientSession(
             *streams) as session:
         initialize_response = await session.initialize()
@@ -86,8 +85,12 @@ def get_tool_description(self,
         pass
 
     @abstractmethod
-    def new_session(self, tool_name: str,
-                    session_id: str) -> AbstractAsyncContextManager[Any]:
+    def new_session(
+        self,
+        tool_name: str,
+        session_id: str,
+        headers: Optional[dict[str, str]] = None
+    ) -> AbstractAsyncContextManager[Any]:
         """
         Create a session for the tool.
         """
@@ -144,16 +147,21 @@ def get_tool_description(self, tool_name: str):
         return self.harmony_tool_descriptions.get(tool_name)
 
     @asynccontextmanager
-    async def new_session(self, tool_name: str, session_id: str):
+    async def new_session(self,
+                          tool_name: str,
+                          session_id: str,
+                          headers: Optional[dict[str, str]] = None):
         from mcp import ClientSession
         from mcp.client.sse import sse_client
         url = self.urls.get(tool_name)
-        headers = {"x-session-id": session_id}
+        request_headers = {"x-session-id": session_id}
+        if headers is not None:
+            request_headers.update(headers)
         if not url:
             raise KeyError(f"Tool '{tool_name}' is not supported")
-        async with sse_client(url=url,
-                              headers=headers) as streams, ClientSession(
-                                  *streams) as session:
+        async with sse_client(
+                url=url, headers=request_headers) as streams, ClientSession(
+                    *streams) as session:
             await session.initialize()
             yield session
 
@@ -189,7 +197,10 @@ def get_tool_description(self,
             raise ValueError(f"Unknown tool {tool_name}")
 
     @asynccontextmanager
-    async def new_session(self, tool_name: str, session_id: str):
+    async def new_session(self,
+                          tool_name: str,
+                          session_id: str,
+                          headers: Optional[dict[str, str]] = None):
         if tool_name not in self.tools:
             raise KeyError(f"Tool '{tool_name}' is not supported")
         yield self.tools[tool_name]