Support v1/responses and use harmony in serving_chat (#8837)

Signed-off-by: Xinyuan Tong <justinning0323@outlook.com>
Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Xinyuan Tong <justinning0323@outlook.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>

python/sglang/srt/entrypoints/openai/protocol.py

@@ -14,9 +14,18 @@
 """Pydantic models for OpenAI API protocol"""
 
 import time
+import uuid
 from dataclasses import dataclass
-from typing import Any, Dict, List, Optional, Union
+from typing import Any, Dict, List, Optional, TypeAlias, Union
 
+from openai.types.responses import (
+    ResponseFunctionToolCall,
+    ResponseInputItemParam,
+    ResponseOutputItem,
+    ResponseReasoningItem,
+)
+from openai.types.responses.response import ToolChoice
+from openai.types.responses.tool import Tool
 from pydantic import (
     BaseModel,
     Field,
@@ -84,6 +93,7 @@ class UsageInfo(BaseModel):
     completion_tokens: Optional[int] = 0
     # only used to return cached tokens when --enable-cache-report is set
     prompt_tokens_details: Optional[Dict[str, int]] = None
+    reasoning_tokens: Optional[int] = 0
 
 
 class StreamOptions(BaseModel):
@@ -428,6 +438,13 @@ class ChatCompletionRequest(BaseModel):
         default="auto", examples=["none"]
     )  # noqa
     return_hidden_states: bool = False
+    reasoning_effort: Optional[Literal["low", "medium", "high"]] = Field(
+        default="medium",
+        description="Constrains effort on reasoning for reasoning models. "
+        "'low' is the least effort, 'high' is the most effort. Reducing reasoning effort can "
+        "result in faster responses and fewer tokens used on reasoning in a response. "
+        "Currently only supported for OpenAI models.",
+    )
 
     @model_validator(mode="before")
     @classmethod
@@ -619,6 +636,196 @@ OpenAIServingRequest = Union[
 ]
 
 
+# Response API protocol definitions
+class ResponseReasoningParam(BaseModel):
+    """Reasoning parameters for responses."""
+
+    effort: Optional[Literal["low", "medium", "high"]] = Field(
+        default="medium",
+        description="Constrains effort on reasoning for reasoning models.",
+    )
+
+
+class ResponseTool(BaseModel):
+    """Tool definition for responses."""
+
+    type: Literal["web_search_preview", "code_interpreter"] = Field(
+        description="Type of tool to enable"
+    )
+
+
+ResponseInputOutputItem: TypeAlias = Union[
+    ResponseInputItemParam,
+    "ResponseReasoningItem",
+    ResponseFunctionToolCall,
+]
+
+
+class ResponsesRequest(BaseModel):
+    """Request body for v1/responses endpoint."""
+
+    # Core OpenAI API fields (ordered by official documentation)
+    background: Optional[bool] = False
+    include: Optional[
+        List[
+            Literal[
+                "code_interpreter_call.outputs",
+                "computer_call_output.output.image_url",
+                "file_search_call.results",
+                "message.input_image.image_url",
+                "message.output_text.logprobs",
+                "reasoning.encrypted_content",
+            ]
+        ]
+    ] = None
+    input: Union[str, List[ResponseInputOutputItem]]
+    instructions: Optional[str] = None
+    max_output_tokens: Optional[int] = None
+    max_tool_calls: Optional[int] = None
+    metadata: Optional[Dict[str, Any]] = None
+    model: Optional[str] = None  # Made optional to match vLLM
+    parallel_tool_calls: Optional[bool] = True
+    previous_response_id: Optional[str] = None
+    reasoning: Optional[ResponseReasoningParam] = None
+    service_tier: Literal["auto", "default", "flex", "scale", "priority"] = "auto"
+    store: Optional[bool] = True
+    stream: Optional[bool] = False
+    temperature: Optional[float] = None
+    tool_choice: Literal["auto", "required", "none"] = "auto"
+    tools: List[ResponseTool] = Field(default_factory=list)
+    top_logprobs: Optional[int] = 0
+    top_p: Optional[float] = None
+    truncation: Optional[Literal["auto", "disabled"]] = "disabled"
+    user: Optional[str] = None
+
+    # Extra SGLang parameters
+    request_id: str = Field(
+        default_factory=lambda: f"resp_{uuid.uuid4().hex}",
+        description="The request_id related to this request. If the caller does not set it, a random uuid will be generated.",
+    )
+    priority: int = Field(default=0, description="Request priority")
+
+    # SGLang-specific sampling parameters
+    frequency_penalty: float = 0.0
+    presence_penalty: float = 0.0
+    stop: Optional[Union[str, List[str]]] = None
+    top_k: int = -1
+    min_p: float = 0.0
+    repetition_penalty: float = 1.0
+
+    # Default sampling parameters
+    _DEFAULT_SAMPLING_PARAMS = {
+        "temperature": 0.7,
+        "top_p": 1.0,
+        "top_k": -1,
+        "min_p": 0.0,
+        "repetition_penalty": 1.0,
+    }
+
+    def to_sampling_params(
+        self, default_max_tokens: int, default_params: Optional[Dict] = None
+    ) -> Dict[str, Any]:
+        """Convert to sampling parameters for generation."""
+        if default_params is None:
+            default_params = {}
+
+        # Use max_output_tokens if available, otherwise use max_tokens for backwards compatibility
+        if self.max_output_tokens is not None:
+            max_tokens = min(self.max_output_tokens, default_max_tokens)
+        else:
+            max_tokens = default_max_tokens
+
+        # Avoid exceed the context length by minus 1 token
+        max_tokens -= 1
+
+        # Get parameters with defaults
+        temperature = self.temperature
+        if temperature is None:
+            temperature = default_params.get(
+                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"]
+            )
+
+        top_p = self.top_p
+        if top_p is None:
+            top_p = default_params.get("top_p", self._DEFAULT_SAMPLING_PARAMS["top_p"])
+
+        params = {
+            "max_new_tokens": max_tokens,
+            "temperature": temperature,
+            "top_p": top_p,
+            "frequency_penalty": self.frequency_penalty,
+            "presence_penalty": self.presence_penalty,
+            "stop": self.stop,
+            "top_k": self.top_k,
+            "min_p": self.min_p,
+            "repetition_penalty": self.repetition_penalty,
+        }
+
+        # Apply any additional default parameters
+        for key, value in default_params.items():
+            if key not in params or params[key] is None:
+                params[key] = value
+
+        return params
+
+
+class PromptTokenUsageInfo(BaseModel):
+    """Prompt token usage details."""
+
+    cached_tokens: int = 0
+
+
+class ResponsesResponse(BaseModel):
+    """Response body for v1/responses endpoint."""
+
+    id: str = Field(default_factory=lambda: f"resp_{time.time()}")
+    object: Literal["response"] = "response"
+    created_at: int = Field(default_factory=lambda: int(time.time()))
+    model: str
+
+    output: List[
+        Union[ResponseOutputItem, ResponseReasoningItem, ResponseFunctionToolCall]
+    ] = Field(default_factory=list)
+    status: Literal["queued", "in_progress", "completed", "failed", "cancelled"]
+    usage: Optional[UsageInfo] = None
+    parallel_tool_calls: bool = True
+    tool_choice: str = "auto"
+    tools: List[ResponseTool] = Field(default_factory=list)
+
+    @classmethod
+    def from_request(
+        cls,
+        request: ResponsesRequest,
+        sampling_params: Any,
+        model_name: str,
+        created_time: int,
+        output: List[
+            Union[ResponseOutputItem, ResponseReasoningItem, ResponseFunctionToolCall]
+        ],
+        status: str,
+        usage: Optional[UsageInfo],
+    ) -> "ResponsesResponse":
+        """Create a response from a request."""
+        return cls(
+            id=request.request_id,
+            created_at=created_time,
+            model=model_name,
+            output=output,
+            status=status,
+            usage=usage,
+            parallel_tool_calls=request.parallel_tool_calls or True,
+            tool_choice=request.tool_choice,
+            tools=request.tools,
+        )
+
+
+class RequestResponseMetadata(BaseModel):
+    """Metadata for request/response tracking."""
+
+    request_id: str
+    final_usage_info: Optional[UsageInfo] = None
+
+
 @dataclass
 class MessageProcessingResult:
     """Result of processing chat messages and applying templates.
@@ -645,3 +852,22 @@ class MessageProcessingResult:
     modalities: List[str]
     stop: List[str]
     tool_call_constraint: Optional[Any] = None
+
+
+class ResponseReasoningTextContent(BaseModel):
+    text: str
+    type: Literal["reasoning_text"] = "reasoning_text"
+
+
+class ResponseReasoningItem(BaseModel):
+    id: str
+    content: list[ResponseReasoningTextContent] = Field(default_factory=list)
+    summary: list = Field(default_factory=list)
+    type: Literal["reasoning"] = "reasoning"
+    encrypted_content: Optional[str] = None
+    status: Optional[Literal["in_progress", "completed", "incomplete"]]
+
+
+ResponseInputOutputItem: TypeAlias = Union[
+    ResponseInputItemParam, "ResponseReasoningItem", ResponseFunctionToolCall
+]

python/sglang/srt/entrypoints/openai/serving_chat.py

@@ -7,8 +7,18 @@ from typing import Any, AsyncGenerator, Dict, List, Optional, Union
 
 from fastapi import Request
 from fastapi.responses import ORJSONResponse, StreamingResponse
+from openai_harmony import Message as OpenAIMessage
 
 from sglang.srt.conversation import generate_chat_conv
+from sglang.srt.entrypoints.harmony_utils import (
+    get_developer_message,
+    get_stop_tokens_for_assistant_actions,
+    get_streamable_parser_for_assistant,
+    get_system_message,
+    parse_chat_input,
+    parse_output_into_messages,
+    render_for_completion,
+)
 from sglang.srt.entrypoints.openai.protocol import (
     ChatCompletionRequest,
     ChatCompletionResponse,
@@ -51,6 +61,26 @@ class OpenAIServingChat(OpenAIServingBase):
     ):
         super().__init__(tokenizer_manager)
         self.template_manager = template_manager
+        self.use_harmony = (
+            self.tokenizer_manager.model_config.hf_config.model_type == "gpt_oss"
+        )
+
+        if self.use_harmony:
+            from sglang.srt.function_call.harmony_tool_parser import (
+                HarmonyToolCallParser,
+            )
+
+            self.harmony_tool_parser = HarmonyToolCallParser()
+
+        # NOTE While OpenAI's chat completion API supports browsing
+        # for some models, currently vLLM doesn't support it. Please use the
+        # Responses API instead.
+        self.supports_browsing = False
+        self.browser_tool = None
+        # NOTE: Chat completion API does not support code interpreter.
+        # Please use the Responses API instead.
+        self.supports_code_interpreter = False
+        self.python_tool = None
 
     def _request_id_prefix(self) -> str:
         return "chatcmpl-"
@@ -77,41 +107,66 @@ class OpenAIServingChat(OpenAIServingBase):
         is_multimodal = self.tokenizer_manager.model_config.is_multimodal
 
         # Process messages and apply chat template
-        processed_messages = self._process_messages(request, is_multimodal)
+        if not self.use_harmony:
+            processed_messages = self._process_messages(request, is_multimodal)
 
-        # Build sampling parameters
-        sampling_params = self._build_sampling_params(
-            request, processed_messages.stop, processed_messages.tool_call_constraint
-        )
+            # Build sampling parameters
+            sampling_params = self._build_sampling_params(
+                request,
+                processed_messages.stop,
+                processed_messages.tool_call_constraint,
+            )
 
-        # Handle single vs multiple requests
-        if is_multimodal:
-            prompt_kwargs = {"text": processed_messages.prompt}
-        else:
-            if isinstance(processed_messages.prompt_ids, str):
-                prompt_kwargs = {"text": processed_messages.prompt_ids}
+            # Handle single vs multiple requests
+            if is_multimodal:
+                prompt_kwargs = {"text": processed_messages.prompt}
             else:
-                prompt_kwargs = {"input_ids": processed_messages.prompt_ids}
+                if isinstance(processed_messages.prompt_ids, str):
+                    prompt_kwargs = {"text": processed_messages.prompt_ids}
+                else:
+                    prompt_kwargs = {"input_ids": processed_messages.prompt_ids}
 
-        adapted_request = GenerateReqInput(
-            **prompt_kwargs,
-            image_data=processed_messages.image_data,
-            video_data=processed_messages.video_data,
-            audio_data=processed_messages.audio_data,
-            sampling_params=sampling_params,
-            return_logprob=request.logprobs,
-            logprob_start_len=-1,
-            top_logprobs_num=request.top_logprobs or 0,
-            stream=request.stream,
-            return_text_in_logprobs=True,
-            modalities=processed_messages.modalities,
-            lora_path=request.lora_path,
-            bootstrap_host=request.bootstrap_host,
-            bootstrap_port=request.bootstrap_port,
-            bootstrap_room=request.bootstrap_room,
-            return_hidden_states=request.return_hidden_states,
-            rid=request.rid,
-        )
+            adapted_request = GenerateReqInput(
+                **prompt_kwargs,
+                image_data=processed_messages.image_data,
+                video_data=processed_messages.video_data,
+                audio_data=processed_messages.audio_data,
+                sampling_params=sampling_params,
+                return_logprob=request.logprobs,
+                logprob_start_len=-1,
+                top_logprobs_num=request.top_logprobs or 0,
+                stream=request.stream,
+                return_text_in_logprobs=True,
+                modalities=processed_messages.modalities,
+                lora_path=request.lora_path,
+                bootstrap_host=request.bootstrap_host,
+                bootstrap_port=request.bootstrap_port,
+                bootstrap_room=request.bootstrap_room,
+                return_hidden_states=request.return_hidden_states,
+                rid=request.rid,
+            )
+        else:
+            processed_messages, prompt_ids = self._make_request_with_harmony(request)
+
+            adapted_request = GenerateReqInput(
+                input_ids=prompt_ids,
+                sampling_params=self._build_sampling_params(
+                    request,
+                    request.stop,
+                    tool_call_constraint=None,
+                ),
+                stream=request.stream,
+                return_logprob=request.logprobs,
+                logprob_start_len=-1,
+                top_logprobs_num=request.top_logprobs or 0,
+                return_text_in_logprobs=True,
+                lora_path=request.lora_path,
+                bootstrap_host=request.bootstrap_host,
+                bootstrap_port=request.bootstrap_port,
+                bootstrap_room=request.bootstrap_room,
+                return_hidden_states=request.return_hidden_states,
+                rid=request.rid,
+            )
 
         return adapted_request, request
 
@@ -402,6 +457,12 @@ class OpenAIServingChat(OpenAIServingBase):
         cached_tokens = {}
         hidden_states = {}
 
+        # Harmony tracking
+        if self.use_harmony:
+            harmony_parsers = [
+                get_streamable_parser_for_assistant() for _ in range(request.n)
+            ]
+
         try:
             async for content in self.tokenizer_manager.generate_request(
                 adapted_request, raw_request
@@ -449,14 +510,57 @@ class OpenAIServingChat(OpenAIServingBase):
                     yield f"data: {chunk.model_dump_json()}\n\n"
 
                 # Process content delta
-                stream_buffer = stream_buffers.get(index, "")
-                delta = content["text"][len(stream_buffer) :]
-                stream_buffers[index] = stream_buffer + delta
+                if self.use_harmony:
+                    harmony_parser = harmony_parsers[index]
+
+                    new_token_ids = content["output_ids"]
+                    for token_id in new_token_ids:
+                        harmony_parser.process(token_id)
+
+                    is_final = harmony_parser.current_channel == "final"
+                    is_analysis = harmony_parser.current_channel == "analysis"
+                    delta = harmony_parser.last_content_delta or ""
+
+                    if is_analysis:
+                        choice_data = ChatCompletionResponseStreamChoice(
+                            index=index,
+                            delta=DeltaMessage(reasoning_content=delta),
+                            finish_reason=None,
+                        )
+                        chunk = ChatCompletionStreamResponse(
+                            id=content["meta_info"]["id"],
+                            created=int(time.time()),
+                            choices=[choice_data],
+                            model=request.model,
+                        )
+                        yield f"data: {chunk.model_dump_json()}\n\n"
+                        continue
+
+                    choice_data = ChatCompletionResponseStreamChoice(
+                        index=index,
+                        delta=DeltaMessage(content=delta if delta else None),
+                        finish_reason=None,
+                        matched_stop=None,
+                        logprobs=choice_logprobs,
+                    )
+                    chunk = ChatCompletionStreamResponse(
+                        id=content["meta_info"]["id"],
+                        created=int(time.time()),
+                        choices=[choice_data],
+                        model=request.model,
+                    )
+                    yield f"data: {chunk.model_dump_json()}\n\n"
+                    continue
+                else:
+                    stream_buffer = stream_buffers.get(index, "")
+                    delta = content["text"][len(stream_buffer) :]
+                    stream_buffers[index] = stream_buffer + delta
 
                 # Handle reasoning content
                 if (
                     self.tokenizer_manager.server_args.reasoning_parser
                     and request.separate_reasoning
+                    and not self.use_harmony
                 ):
                     reasoning_text, delta = self._process_reasoning_stream(
                         index, delta, reasoning_parser_dict, content, request
@@ -475,8 +579,27 @@ class OpenAIServingChat(OpenAIServingBase):
                         )
                         yield f"data: {chunk.model_dump_json()}\n\n"
 
+                if self.use_harmony and not is_final:
+                    choice_data = ChatCompletionResponseStreamChoice(
+                        index=index,
+                        delta=DeltaMessage(reasoning_content=delta),
+                        finish_reason=None,
+                    )
+                    chunk = ChatCompletionStreamResponse(
+                        id=content["meta_info"]["id"],
+                        created=int(time.time()),
+                        choices=[choice_data],
+                        model=request.model,
+                    )
+                    yield f"data: {chunk.model_dump_json()}\n\n"
+
                 # Handle tool calls
-                if request.tool_choice != "none" and request.tools:
+                # TODO: support tool call parsing for harmony
+                if (
+                    request.tool_choice != "none"
+                    and request.tools
+                    and not self.use_harmony
+                ):
                     async for chunk in self._process_tool_call_stream(
                         index,
                         delta,
@@ -502,7 +625,7 @@ class OpenAIServingChat(OpenAIServingBase):
                     if delta:
                         choice_data = ChatCompletionResponseStreamChoice(
                             index=index,
-                            delta=DeltaMessage(content=delta if delta else None),
+                            delta=DeltaMessage(content=delta),
                             finish_reason=None,
                             matched_stop=None,
                             logprobs=choice_logprobs,
@@ -640,6 +763,76 @@ class OpenAIServingChat(OpenAIServingBase):
 
             finish_reason = ret_item["meta_info"]["finish_reason"]
             text = ret_item["text"]
+            output_ids = ret_item["output_ids"]
+
+            if self.use_harmony:
+                parser = parse_output_into_messages(output_ids)
+                output_msgs = parser.messages
+                if len(output_msgs) == 0:
+                    # The generation has stopped during reasoning.
+                    is_tool_call = False
+                    reasoning_content = parser.current_content
+                    final_content = None
+                elif len(output_msgs) == 1:
+                    # The generation has stopped during final message.
+                    is_tool_call = False
+                    reasoning_content = output_msgs[0].content[0].text
+                    final_content = parser.current_content
+                else:
+                    if len(output_msgs) != 2:
+                        raise ValueError(
+                            "Expected 2 output messages (reasoning and final), "
+                            f"but got {len(output_msgs)}."
+                        )
+                    reasoning_msg, final_msg = output_msgs
+                    reasoning_content = reasoning_msg.content[0].text
+                    final_content = final_msg.content[0].text
+                    is_tool_call = final_msg.recipient is not None
+
+                if is_tool_call:
+                    # Extract tool call information from final message
+                    tool_call = (
+                        self.harmony_tool_parser.extract_tool_calls_from_message(
+                            final_msg
+                        )
+                    )
+                    tool_calls = [tool_call] if tool_call else []
+
+                    message = ChatMessage(
+                        role="assistant",
+                        reasoning_content=reasoning_content,
+                        content=None,  # Tool calls don't have regular content
+                        tool_calls=tool_calls,
+                    )
+                else:
+                    # Normal message
+                    message = ChatMessage(
+                        role="assistant",
+                        reasoning_content=reasoning_content,
+                        content=final_content,
+                    )
+
+                if is_tool_call:
+                    finish_reason_type = "tool_calls"
+                elif finish_reason:
+                    finish_reason_type = (
+                        finish_reason["type"] if finish_reason else "stop"
+                    )
+                else:
+                    finish_reason_type = "stop"
+                choice_data = ChatCompletionResponseChoice(
+                    index=idx,
+                    message=message,
+                    logprobs=choice_logprobs,
+                    finish_reason=finish_reason_type,
+                    matched_stop=(
+                        finish_reason["matched"]
+                        if finish_reason and "matched" in finish_reason
+                        else None
+                    ),
+                )
+                choices.append(choice_data)
+                continue
 
             # Handle reasoning content
             reasoning_text = None
@@ -978,3 +1171,33 @@ class OpenAIServingChat(OpenAIServingBase):
             return f"data: {chunk.model_dump_json()}\n\n"
 
         return None
+
+    def _make_request_with_harmony(
+        self,
+        request: ChatCompletionRequest,
+    ):
+        messages: list[OpenAIMessage] = []
+
+        # Add system message.
+        # In Chat Completion API, browsing is enabled by default if the model
+        # supports it.
+        assert not self.supports_browsing
+        assert not self.supports_code_interpreter
+        sys_msg = get_system_message(
+            reasoning_effort=request.reasoning_effort,
+            browser_description=None,
+            python_description=None,
+        )
+        messages.append(sys_msg)
+
+        # Add developer message.
+        dev_msg = get_developer_message()
+        messages.append(dev_msg)
+
+        # Add user message.
+        for chat_msg in request.messages:
+            messages.append(parse_chat_input(chat_msg))
+
+        # Render prompt token ids.
+        prompt_token_ids = render_for_completion(messages)
+        return messages, prompt_token_ids

python/sglang/srt/entrypoints/openai/tool_server.py

@@ -0,0 +1,174 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import logging
+from abc import ABC, abstractmethod
+from contextlib import AbstractAsyncContextManager, asynccontextmanager
+from typing import Any
+
+logger = logging.getLogger(__name__)
+try:
+    from mcp import ClientSession
+    from mcp.client.sse import sse_client
+    from mcp.types import ListToolsResult
+except ImportError:
+    logger.warning("Ignoring mcp import error")
+
+from openai_harmony import ToolDescription, ToolNamespaceConfig
+
+
+async def list_server_and_tools(server_url: str):
+
+    async with sse_client(url=server_url) as streams, ClientSession(
+        *streams
+    ) as session:
+        initialize_response = await session.initialize()
+        list_tools_response = await session.list_tools()
+        return initialize_response, list_tools_response
+
+
+def trim_schema(schema: dict) -> dict:
+    # Turn JSON Schema from MCP generated into Harmony's variant.
+    if "title" in schema:
+        del schema["title"]
+    if "default" in schema and schema["default"] is None:
+        del schema["default"]
+    if "anyOf" in schema:
+        # Turn "anyOf": [{"type": "type-1"}, {"type": "type-2"}]
+        # into "type": ["type-1", "type-2"]
+        # if there's more than 1 types, also remove "null" type as Harmony will
+        # just ignore it
+        types = [
+            type_dict["type"]
+            for type_dict in schema["anyOf"]
+            if type_dict["type"] != "null"
+        ]
+        schema["type"] = types
+        del schema["anyOf"]
+    if "properties" in schema:
+        schema["properties"] = {
+            k: trim_schema(v) for k, v in schema["properties"].items()
+        }
+    return schema
+
+
+def post_process_tools_description(
+    list_tools_result: "ListToolsResult",
+) -> "ListToolsResult":
+    # Adapt the MCP tool result for Harmony
+    for tool in list_tools_result.tools:
+        tool.inputSchema = trim_schema(tool.inputSchema)
+
+    # Some tools schema don't need to be part of the prompt (e.g. simple text
+    # in text out for Python)
+    list_tools_result.tools = [
+        tool
+        for tool in list_tools_result.tools
+        if getattr(tool.annotations, "include_in_prompt", True)
+    ]
+
+    return list_tools_result
+
+
+class ToolServer(ABC):
+
+    @abstractmethod
+    def has_tool(self, tool_name: str):
+        pass
+
+    @abstractmethod
+    def get_tool_description(self, tool_name: str):
+        pass
+
+    @abstractmethod
+    def get_tool_session(self, tool_name: str) -> AbstractAsyncContextManager[Any]: ...
+
+
+class MCPToolServer(ToolServer):
+
+    def __init__(self):
+        self.harmony_tool_descriptions = {}
+
+    async def add_tool_server(self, server_url: str):
+        tool_urls = server_url.split(",")
+        self.harmony_tool_descriptions = {}
+        self.urls: dict[str, str] = {}
+        for url in tool_urls:
+            url = f"http://{url}/sse"
+            initialize_response, list_tools_response = await list_server_and_tools(url)
+
+            list_tools_response = post_process_tools_description(list_tools_response)
+
+            tool_from_mcp = ToolNamespaceConfig(
+                name=initialize_response.serverInfo.name,
+                description=initialize_response.instructions,
+                tools=[
+                    ToolDescription.new(
+                        name=tool.name,
+                        description=tool.description,
+                        parameters=tool.inputSchema,
+                    )
+                    for tool in list_tools_response.tools
+                ],
+            )
+            self.harmony_tool_descriptions[tool_from_mcp.name] = tool_from_mcp
+            if tool_from_mcp.name not in self.urls:
+                self.urls[tool_from_mcp.name] = url
+            else:
+                logger.warning(
+                    "Tool %s already exists. Ignoring duplicate tool server %s",
+                    tool_from_mcp.name,
+                    url,
+                )
+
+    def has_tool(self, tool_name: str):
+        return tool_name in self.harmony_tool_descriptions
+
+    def get_tool_description(self, tool_name: str):
+        return self.harmony_tool_descriptions.get(tool_name)
+
+    @asynccontextmanager
+    async def get_tool_session(self, tool_name: str):
+        url = self.urls.get(tool_name)
+        if url:
+            async with sse_client(url=url) as streams, ClientSession(
+                *streams
+            ) as session:
+                await session.initialize()
+                yield session
+        else:
+            logger.warning("Tool %s not found", tool_name)
+
+
+class DemoToolServer(ToolServer):
+
+    def __init__(self):
+        from sglang.srt.entrypoints.tool import (
+            HarmonyBrowserTool,
+            HarmonyPythonTool,
+            Tool,
+        )
+
+        self.tools: dict[str, Tool] = {}
+        browser_tool = HarmonyBrowserTool()
+        if browser_tool.enabled:
+            self.tools["browser"] = browser_tool
+        python_tool = HarmonyPythonTool()
+        if python_tool.enabled:
+            self.tools["python"] = python_tool
+
+    def has_tool(self, tool_name: str):
+        return tool_name in self.tools
+
+    def get_tool_description(self, tool_name: str):
+        if tool_name not in self.tools:
+            return None
+        if tool_name == "browser":
+            return ToolNamespaceConfig.browser()
+        elif tool_name == "python":
+            return ToolNamespaceConfig.python()
+        else:
+            raise ValueError(f"Unknown tool {tool_name}")
+
+    @asynccontextmanager
+    async def get_tool_session(self, tool_name: str):
+        yield self.tools[tool_name]