import json import logging from contextlib import AsyncExitStack from datetime import timedelta from pathlib import Path from typing import TYPE_CHECKING, Any, AsyncIterable, Dict, List, Literal, Optional, Union, overload from typing_extensions import NotRequired, TypeAlias, TypedDict, Unpack from ...utils._runtime import get_hf_hub_version from .._generated._async_client import AsyncInferenceClient from .._generated.types import ( ChatCompletionInputMessage, ChatCompletionInputTool, ChatCompletionStreamOutput, ChatCompletionStreamOutputDeltaToolCall, ) from .._providers import PROVIDER_OR_POLICY_T from .utils import format_result if TYPE_CHECKING: from mcp import ClientSession logger = logging.getLogger(__name__) # Type alias for tool names ToolName: TypeAlias = str ServerType: TypeAlias = Literal["stdio", "sse", "http"] class StdioServerParameters_T(TypedDict): command: str args: NotRequired[List[str]] env: NotRequired[Dict[str, str]] cwd: NotRequired[Union[str, Path, None]] class SSEServerParameters_T(TypedDict): url: str headers: NotRequired[Dict[str, Any]] timeout: NotRequired[float] sse_read_timeout: NotRequired[float] class StreamableHTTPParameters_T(TypedDict): url: str headers: NotRequired[dict[str, Any]] timeout: NotRequired[timedelta] sse_read_timeout: NotRequired[timedelta] terminate_on_close: NotRequired[bool] class MCPClient: """ Client for connecting to one or more MCP servers and processing chat completions with tools. > [!WARNING] > This class is experimental and might be subject to breaking changes in the future without prior notice. Args: model (`str`, `optional`): The model to run inference with. Can be a model id hosted on the Hugging Face Hub, e.g. `meta-llama/Meta-Llama-3-8B-Instruct` or a URL to a deployed Inference Endpoint or other local or remote endpoint. provider (`str`, *optional*): Name of the provider to use for inference. Defaults to "auto" i.e. the first of the providers available for the model, sorted by the user's order in https://hf.co/settings/inference-providers. If model is a URL or `base_url` is passed, then `provider` is not used. base_url (`str`, *optional*): The base URL to run inference. Defaults to None. api_key (`str`, `optional`): Token to use for authentication. Will default to the locally Hugging Face saved token if not provided. You can also use your own provider API key to interact directly with the provider's service. """ def __init__( self, *, model: Optional[str] = None, provider: Optional[PROVIDER_OR_POLICY_T] = None, base_url: Optional[str] = None, api_key: Optional[str] = None, ): # Initialize MCP sessions as a dictionary of ClientSession objects self.sessions: Dict[ToolName, "ClientSession"] = {} self.exit_stack = AsyncExitStack() self.available_tools: List[ChatCompletionInputTool] = [] # To be able to send the model in the payload if `base_url` is provided if model is None and base_url is None: raise ValueError("At least one of `model` or `base_url` should be set in `MCPClient`.") self.payload_model = model self.client = AsyncInferenceClient( model=None if base_url is not None else model, provider=provider, api_key=api_key, base_url=base_url, ) async def __aenter__(self): """Enter the context manager""" await self.client.__aenter__() await self.exit_stack.__aenter__() return self async def __aexit__(self, exc_type, exc_val, exc_tb): """Exit the context manager""" await self.client.__aexit__(exc_type, exc_val, exc_tb) await self.cleanup() async def cleanup(self): """Clean up resources""" await self.client.close() await self.exit_stack.aclose() @overload async def add_mcp_server(self, type: Literal["stdio"], **params: Unpack[StdioServerParameters_T]): ... @overload async def add_mcp_server(self, type: Literal["sse"], **params: Unpack[SSEServerParameters_T]): ... @overload async def add_mcp_server(self, type: Literal["http"], **params: Unpack[StreamableHTTPParameters_T]): ... async def add_mcp_server(self, type: ServerType, **params: Any): """Connect to an MCP server Args: type (`str`): Type of the server to connect to. Can be one of: - "stdio": Standard input/output server (local) - "sse": Server-sent events (SSE) server - "http": StreamableHTTP server **params (`Dict[str, Any]`): Server parameters that can be either: - For stdio servers: - command (str): The command to run the MCP server - args (List[str], optional): Arguments for the command - env (Dict[str, str], optional): Environment variables for the command - cwd (Union[str, Path, None], optional): Working directory for the command - allowed_tools (List[str], optional): List of tool names to allow from this server - For SSE servers: - url (str): The URL of the SSE server - headers (Dict[str, Any], optional): Headers for the SSE connection - timeout (float, optional): Connection timeout - sse_read_timeout (float, optional): SSE read timeout - allowed_tools (List[str], optional): List of tool names to allow from this server - For StreamableHTTP servers: - url (str): The URL of the StreamableHTTP server - headers (Dict[str, Any], optional): Headers for the StreamableHTTP connection - timeout (timedelta, optional): Connection timeout - sse_read_timeout (timedelta, optional): SSE read timeout - terminate_on_close (bool, optional): Whether to terminate on close - allowed_tools (List[str], optional): List of tool names to allow from this server """ from mcp import ClientSession, StdioServerParameters from mcp import types as mcp_types # Extract allowed_tools configuration if provided allowed_tools = params.pop("allowed_tools", None) # Determine server type and create appropriate parameters if type == "stdio": # Handle stdio server from mcp.client.stdio import stdio_client logger.info(f"Connecting to stdio MCP server with command: {params['command']} {params.get('args', [])}") client_kwargs = {"command": params["command"]} for key in ["args", "env", "cwd"]: if params.get(key) is not None: client_kwargs[key] = params[key] server_params = StdioServerParameters(**client_kwargs) read, write = await self.exit_stack.enter_async_context(stdio_client(server_params)) elif type == "sse": # Handle SSE server from mcp.client.sse import sse_client logger.info(f"Connecting to SSE MCP server at: {params['url']}") client_kwargs = {"url": params["url"]} for key in ["headers", "timeout", "sse_read_timeout"]: if params.get(key) is not None: client_kwargs[key] = params[key] read, write = await self.exit_stack.enter_async_context(sse_client(**client_kwargs)) elif type == "http": # Handle StreamableHTTP server from mcp.client.streamable_http import streamablehttp_client logger.info(f"Connecting to StreamableHTTP MCP server at: {params['url']}") client_kwargs = {"url": params["url"]} for key in ["headers", "timeout", "sse_read_timeout", "terminate_on_close"]: if params.get(key) is not None: client_kwargs[key] = params[key] read, write, _ = await self.exit_stack.enter_async_context(streamablehttp_client(**client_kwargs)) # ^ TODO: should be handle `get_session_id_callback`? (function to retrieve the current session ID) else: raise ValueError(f"Unsupported server type: {type}") session = await self.exit_stack.enter_async_context( ClientSession( read_stream=read, write_stream=write, client_info=mcp_types.Implementation( name="huggingface_hub.MCPClient", version=get_hf_hub_version(), ), ) ) logger.debug("Initializing session...") await session.initialize() # List available tools response = await session.list_tools() logger.debug("Connected to server with tools:", [tool.name for tool in response.tools]) # Filter tools based on allowed_tools configuration filtered_tools = response.tools if allowed_tools is not None: filtered_tools = [tool for tool in response.tools if tool.name in allowed_tools] logger.debug( f"Tool filtering applied. Using {len(filtered_tools)} of {len(response.tools)} available tools: {[tool.name for tool in filtered_tools]}" ) for tool in filtered_tools: if tool.name in self.sessions: logger.warning(f"Tool '{tool.name}' already defined by another server. Skipping.") continue # Map tool names to their server for later lookup self.sessions[tool.name] = session # Add tool to the list of available tools (for use in chat completions) self.available_tools.append( ChatCompletionInputTool.parse_obj_as_instance( { "type": "function", "function": { "name": tool.name, "description": tool.description, "parameters": tool.inputSchema, }, } ) ) async def process_single_turn_with_tools( self, messages: List[Union[Dict, ChatCompletionInputMessage]], exit_loop_tools: Optional[List[ChatCompletionInputTool]] = None, exit_if_first_chunk_no_tool: bool = False, ) -> AsyncIterable[Union[ChatCompletionStreamOutput, ChatCompletionInputMessage]]: """Process a query using `self.model` and available tools, yielding chunks and tool outputs. Args: messages (`List[Dict]`): List of message objects representing the conversation history exit_loop_tools (`List[ChatCompletionInputTool]`, *optional*): List of tools that should exit the generator when called exit_if_first_chunk_no_tool (`bool`, *optional*): Exit if no tool is present in the first chunks. Default to False. Yields: [`ChatCompletionStreamOutput`] chunks or [`ChatCompletionInputMessage`] objects """ # Prepare tools list based on options tools = self.available_tools if exit_loop_tools is not None: tools = [*exit_loop_tools, *self.available_tools] # Create the streaming request response = await self.client.chat.completions.create( model=self.payload_model, messages=messages, tools=tools, tool_choice="auto", stream=True, ) message: Dict[str, Any] = {"role": "unknown", "content": ""} final_tool_calls: Dict[int, ChatCompletionStreamOutputDeltaToolCall] = {} num_of_chunks = 0 # Read from stream async for chunk in response: num_of_chunks += 1 delta = chunk.choices[0].delta if chunk.choices and len(chunk.choices) > 0 else None if not delta: continue # Process message if delta.role: message["role"] = delta.role if delta.content: message["content"] += delta.content # Process tool calls if delta.tool_calls: for tool_call in delta.tool_calls: idx = tool_call.index # first chunk for this tool call if idx not in final_tool_calls: final_tool_calls[idx] = tool_call if final_tool_calls[idx].function.arguments is None: final_tool_calls[idx].function.arguments = "" continue # safety before concatenating text to .function.arguments if final_tool_calls[idx].function.arguments is None: final_tool_calls[idx].function.arguments = "" if tool_call.function.arguments: final_tool_calls[idx].function.arguments += tool_call.function.arguments # Optionally exit early if no tools in first chunks if exit_if_first_chunk_no_tool and num_of_chunks <= 2 and len(final_tool_calls) == 0: return # Yield each chunk to caller yield chunk # Add the assistant message with tool calls (if any) to messages if message["content"] or final_tool_calls: # if the role is unknown, set it to assistant if message.get("role") == "unknown": message["role"] = "assistant" # Convert final_tool_calls to the format expected by OpenAI if final_tool_calls: tool_calls_list: List[Dict[str, Any]] = [] for tc in final_tool_calls.values(): tool_calls_list.append( { "id": tc.id, "type": "function", "function": { "name": tc.function.name, "arguments": tc.function.arguments or "{}", }, } ) message["tool_calls"] = tool_calls_list messages.append(message) # Process tool calls one by one for tool_call in final_tool_calls.values(): function_name = tool_call.function.name try: function_args = json.loads(tool_call.function.arguments or "{}") except json.JSONDecodeError as err: tool_message = { "role": "tool", "tool_call_id": tool_call.id, "name": function_name, "content": f"Invalid JSON generated by the model: {err}", } tool_message_as_obj = ChatCompletionInputMessage.parse_obj_as_instance(tool_message) messages.append(tool_message_as_obj) yield tool_message_as_obj continue # move to next tool call tool_message = {"role": "tool", "tool_call_id": tool_call.id, "content": "", "name": function_name} # Check if this is an exit loop tool if exit_loop_tools and function_name in [t.function.name for t in exit_loop_tools]: tool_message_as_obj = ChatCompletionInputMessage.parse_obj_as_instance(tool_message) messages.append(tool_message_as_obj) yield tool_message_as_obj return # Execute tool call with the appropriate session session = self.sessions.get(function_name) if session is not None: try: result = await session.call_tool(function_name, function_args) tool_message["content"] = format_result(result) except Exception as err: tool_message["content"] = f"Error: MCP tool call failed with error message: {err}" else: tool_message["content"] = f"Error: No session found for tool: {function_name}" # Yield tool message tool_message_as_obj = ChatCompletionInputMessage.parse_obj_as_instance(tool_message) messages.append(tool_message_as_obj) yield tool_message_as_obj