mcp_openai_client.py

import asyncio
import os
import json
from typing import List, Dict, Any, Union
from contextlib import AsyncExitStack
from datetime import datetime
import gradio as gr
from gradio.components.chatbot import ChatMessage
from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client
from mcp.client.sse import sse_client
from anthropic import Anthropic
from anthropic._exceptions import OverloadedError
from dotenv import load_dotenv
from openai import OpenAI
import openai
from openai.types.responses import (
    ResponseTextDeltaEvent,
    ResponseContentPartAddedEvent,
    ResponseContentPartDoneEvent,
    ResponseTextDoneEvent,
    ResponseMcpCallInProgressEvent,
    ResponseAudioDeltaEvent,
    ResponseMcpCallCompletedEvent,
    ResponseOutputItemDoneEvent,
    ResponseOutputItemAddedEvent,
    ResponseCompletedEvent,
)
import ast

load_dotenv()

# LLM_PROVIDER = "anthropic"
LLM_PROVIDER = "openai"

SYSTEM_PROMPT = f"""You are a helpful assistant. Today is {datetime.now().strftime("%Y-%m-%d")}.

You **do not** have prior knowledge of the World Development Indicators (WDI) data. Instead, you must rely entirely on the tools available to you to answer the user's questions.

Detect the language of the user's query and use that language for your response, unless the user specifies otherwise.

When responding you must always plan the steps and enumerate all the tools that you plan to use to answer the user's query.

### Your Instructions:

1. **Tool Use Only**:
   - You must not provide any answers based on prior knowledge or assumptions.
   - You must **not** fabricate data or simulate the behavior of the `get_wdi_data` tool.
   - You cannot use the `get_wdi_data` tool without using the `search_relevant_indicators` tool first.
   - If the user requests WDI data, you **MUST ALWAYS** first call the `search_relevant_indicators` tool to see if there's any relevant data.
   - If relevant data exists, call the `get_wdi_data` tool to get the data.

2. **Tool Invocation**:
   - Use any relevant tools provided to you to answer the user's question.
   - You may call multiple tools if needed, and you should do so in a logical sequence to minimize unnecessary user interaction.
   - Do not hesitate to invoke tools as soon as they are relevant.

3. **Limitations**:
   - If a user request cannot be fulfilled using the tools available, respond by clearly stating that you do not have access to that information.

4. **Ethical Guidelines**:
   - Do not make or endorse statements based on stereotypes, bias, or assumptions.
   - Ensure all claims and explanations are grounded in the data or factual evidence retrieved via tools.
   - Politely refuse to respond to requests that involve stereotypes or unfounded generalizations.

5. **Communication Style**:
   - Present the data in clear, user-friendly language.
   - You may summarize or explain the data retrieved, but do **not** elaborate based on outside or implicit knowledge.
   - You may describe the data in a way that is easy to understand but you MUST NOT elaborate based on external knowledge.
   - Provide summary of the answer in the last step describing some observations and insights solely based on the data.

6. **Presentation**:
   - Present the data in a way that is easy to understand.
   - Summarize the data in a table format with clear column names and values.
   - If the data is not available, respond by clearly stating that you do not have access to that information.

7. **Tool Use**:
   - Fetch each indicator data using independent tool calls.
   - Provide some brief explanation between tool calls.

Stay strictly within these boundaries while maintaining a helpful and respectful tone."""


LLM_MODEL = "claude-3-5-haiku-20241022"
OPENAI_MODEL = "gpt-4.1"
# OPENAI_MODEL = "gpt-4.1-mini"
# OPENAI_MODEL = "gpt-4.1-nano"
# What is the military spending of bangladesh in 2014?
# When a tool is needed for any step, ensure to add the token `TOOL_USE`.


loop = asyncio.new_event_loop()
asyncio.set_event_loop(loop)


class MCPClientWrapper:
    def __init__(self):
        self.session = None
        self.exit_stack = None
        self.anthropic = Anthropic()
        self.openai = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))
        self.tools = []

    async def connect(self, server_path_or_url: str) -> str:
        try:
            # If there's an existing session, close it
            if self.exit_stack:
                return "Already connected to an MCP server. Please disconnect first."
                # await self.exit_stack.aclose()

            self.exit_stack = AsyncExitStack()

            if server_path_or_url.endswith(".py"):
                command = "python"

                server_params = StdioServerParameters(
                    command=command,
                    args=[server_path_or_url],
                    env={"PYTHONIOENCODING": "utf-8", "PYTHONUNBUFFERED": "1"},
                )

                print(
                    f"Starting MCP server with command: {command} {server_path_or_url}"
                )
                # Launch MCP subprocess and bind streams on the current running loop
                stdio_transport = await self.exit_stack.enter_async_context(
                    stdio_client(server_params)
                )
                self.stdio, self.write = stdio_transport
            else:
                print(f"Connecting to MCP server at: {server_path_or_url}")
                sse_transport = await self.exit_stack.enter_async_context(
                    sse_client(
                        server_path_or_url,
                        headers={"Authorization": f"Bearer {os.getenv('HF_TOKEN')}"},
                    )
                )
                self.stdio, self.write = sse_transport

            print("Creating MCP client session...")
            # Create ClientSession on this same loop
            self.session = await self.exit_stack.enter_async_context(
                ClientSession(self.stdio, self.write)
            )
            await self.session.initialize()
            print("MCP session initialized successfully")

            response = await self.session.list_tools()
            self.tools = [
                {
                    "name": tool.name,
                    "description": tool.description,
                    "input_schema": tool.inputSchema,
                }
                for tool in response.tools
            ]

            print("Available tools:", self.tools)
            tool_names = [tool["name"] for tool in self.tools]
            return f"Connected to MCP server. Available tools: {', '.join(tool_names)}"
        except Exception as e:
            error_msg = f"Failed to connect to MCP server: {str(e)}"
            print(error_msg)
            # Clean up on error
            if self.exit_stack:
                await self.exit_stack.aclose()
                self.exit_stack = None
            self.session = None
            return error_msg

    async def disconnect(self):
        if self.exit_stack:
            print("Disconnecting from MCP server...")
            await self.exit_stack.aclose()
            self.exit_stack = None
            self.session = None

    async def process_message(
        self,
        message: str,
        history: List[Union[Dict[str, Any], ChatMessage]],
        previous_response_id: str = None,
    ):
        print("previous_response_id", previous_response_id)
        if not self.session and LLM_PROVIDER == "anthropic":
            messages = history + [
                {"role": "user", "content": message},
                {
                    "role": "assistant",
                    "content": "Please connect to an MCP server first by reloading the page.",
                },
            ]
            yield messages, gr.Textbox(value=""), previous_response_id
        else:
            messages = history + [
                {"role": "user", "content": message},
                {
                    "role": "assistant",
                    "content": "Ok, let me think about your query 🤔...",
                },
            ]

            yield messages, gr.Textbox(value=""), previous_response_id
            # simulate thinking with asyncio.sleep
            await asyncio.sleep(0.2)
            messages.pop(-1)

            is_delta = False
            async for partial in self._process_query(
                message, history, previous_response_id
            ):
                if partial[-1].get("delta"):
                    if not is_delta:
                        is_delta = True
                        messages.append(
                            {
                                "role": "assistant",
                                "content": "",
                            }
                        )
                    messages[-1]["content"] += partial[-1]["delta"]
                    if partial[-1].get("status") == "done":
                        await asyncio.sleep(0.05)
                else:
                    is_delta = False
                    if partial[-1].get("response_id"):
                        previous_response_id = partial[-1]["response_id"]
                        yield (
                            messages,
                            gr.Textbox(value=""),
                            previous_response_id,
                        )
                        await asyncio.sleep(0.01)
                        continue
                    else:
                        messages.extend(partial)
                    print(partial)

                yield (
                    messages,
                    gr.Textbox(value=""),
                    previous_response_id,
                )
                await asyncio.sleep(0.01)

                if (
                    messages[-1]["role"] == "assistant"
                    and messages[-1]["content"]
                    == "The LLM API is overloaded now, try again later..."
                ):
                    break

        with open("messages.log.jsonl", "a+") as fl:
            fl.write(
                json.dumps(
                    dict(
                        time=f"{datetime.now()}",
                        messages=messages,
                        previous_response_id=previous_response_id,
                    )
                )
            )

    async def _process_query_openai(
        self,
        message: str,
        history: List[Union[Dict[str, Any], ChatMessage]],
        previous_response_id: str = None,
    ):
        response = self.openai.responses.create(
            model=OPENAI_MODEL,
            tools=[
                {
                    "type": "mcp",
                    "server_label": "wdi_mcp",
                    "server_url": "https://avsolatorio-test-data-mcp-server.hf.space/gradio_api/mcp/sse",
                    "require_approval": "never",
                    "headers": {"Authorization": f"Bearer {os.getenv('HF_TOKEN')}"},
                    # "server_token": userdata.get('MCP_HF_TOKEN'),
                },
            ],
            # input="What transport protocols are supported in the 2025-03-26 version of the MCP spec?",
            instructions=SYSTEM_PROMPT,
            # input="What is the gdp of india in 2020?",
            input=message,
            parallel_tool_calls=False,
            stream=True,
            max_output_tokens=32768,
            temperature=0,
            previous_response_id=previous_response_id.strip()
            if previous_response_id
            else None,
            store=True,  # Store the response in the OpenAIlogs
        )

        is_tool_call = False
        tool_name = None
        tool_args = None
        for event in response:
            if isinstance(event, ResponseCompletedEvent):
                yield [
                    {
                        "response_id": event.response.id,
                    }
                ]
            elif (
                isinstance(event, ResponseOutputItemAddedEvent)
                and event.item.type == "mcp_call"
            ):
                is_tool_call = True
                tool_name = event.item.name
            # if isinstance(event, ResponseMcpCallInProgressEvent):
            #     is_tool_call = True
            #     yield [
            #         {
            #             "role": "assistant",
            #             "content": "I'll use the tool to help answer your question.",
            #         }
            #     ]
            if is_tool_call:
                if (
                    isinstance(event, ResponseAudioDeltaEvent)
                    and event.type == "response.mcp_call_arguments.done"
                ):
                    tool_args = event.arguments

                    try:
                        tool_args = json.dumps(
                            json.loads(tool_args), ensure_ascii=True, indent=2
                        )
                    except:
                        pass

                    yield [
                        {
                            "role": "assistant",
                            "content": f"I'll use the {tool_name} tool to help answer your question.",
                            "metadata": {
                                "title": f"Using tool: {tool_name.replace('avsolatorio_test_data_mcp_server', '')}",
                                "log": f"Parameters: {tool_args}",
                                # "status": "pending",
                                "status": "done",
                                "id": f"tool_call_{tool_name}",
                            },
                        }
                    ]

                    yield [
                        {
                            "role": "assistant",
                            "content": "```json\n" + tool_args + "\n```",
                            "metadata": {
                                "parent_id": f"tool_call_{tool_name}",
                                "id": f"params_{tool_name}",
                                "title": "Tool Parameters",
                            },
                        }
                    ]

                elif isinstance(event, ResponseOutputItemDoneEvent):
                    if event.item.type == "mcp_call":
                        yield [
                            {
                                "role": "assistant",
                                "content": "Here are the results from the tool:",
                                "metadata": {
                                    "title": f"Tool Result for {tool_name.replace('avsolatorio_test_data_mcp_server', '')}",
                                    "status": "done",
                                    "id": f"result_{tool_name}",
                                },
                            }
                        ]

                        result_content = event.item.output
                        if result_content.startswith("root="):
                            result_content = result_content[5:]
                            try:
                                result_content = ast.literal_eval(result_content)
                                result_content = json.dumps(result_content, indent=2)
                            except:
                                pass

                        yield [
                            {
                                "role": "assistant",
                                "content": "```\n" + result_content + "\n```",
                                "metadata": {
                                    "parent_id": f"result_{tool_name}",
                                    "id": f"raw_result_{tool_name}",
                                    "title": "Raw Output",
                                },
                            }
                        ]
                        is_tool_call = False
                        tool_name = None
                        tool_args = None

            elif (
                isinstance(event, ResponseContentPartDoneEvent)
                and event.type == "response.content_part.done"
            ):
                yield [
                    {
                        "role": "assistant",
                        "content": "",
                        "delta": "",
                        "status": "done",
                    }
                ]
            elif isinstance(event, ResponseTextDeltaEvent):
                yield [{"role": "assistant", "content": None, "delta": event.delta}]

    async def _process_query_anthropic(
        self, message: str, history: List[Union[Dict[str, Any], ChatMessage]]
    ):
        claude_messages = []
        for msg in history:
            if isinstance(msg, ChatMessage):
                role, content = msg.role, msg.content
            else:
                role, content = msg.get("role"), msg.get("content")

            if role in ["user", "assistant", "system"]:
                claude_messages.append({"role": role, "content": content})

        claude_messages.append({"role": "user", "content": message})

        try:
            response = self.anthropic.messages.create(
                # model="claude-3-5-sonnet-20241022",
                model=LLM_MODEL,
                system=SYSTEM_PROMPT,
                max_tokens=1000,
                messages=claude_messages,
                tools=self.tools,
            )
        except OverloadedError:
            yield [
                {
                    "role": "assistant",
                    "content": "The LLM API is overloaded now, try again later...",
                }
            ]
            # TODO: Add a retry mechanism

        result_messages = []
        partial_messages = []

        print(response.content)
        contents = response.content

        MAX_CALLS = 10
        auto_calls = 0

        while len(contents) > 0 and auto_calls < MAX_CALLS:
            content = contents.pop(0)

            if content.type == "text":
                result_messages.append({"role": "assistant", "content": content.text})
                claude_messages.append({"role": "assistant", "content": content.text})
                partial_messages.append(result_messages[-1])
                yield [result_messages[-1]]
                partial_messages = []

            elif content.type == "tool_use":
                tool_id = content.id
                tool_name = content.name
                tool_args = content.input

                result_messages.append(
                    {
                        "role": "assistant",
                        "content": f"I'll use the {tool_name} tool to help answer your question.",
                        "metadata": {
                            "title": f"Using tool: {tool_name.replace('avsolatorio_test_data_mcp_server', '')}",
                            "log": f"Parameters: {json.dumps(tool_args, ensure_ascii=True)}",
                            # "status": "pending",
                            "status": "done",
                            "id": f"tool_call_{tool_name}",
                        },
                    }
                )
                partial_messages.append(result_messages[-1])
                yield [result_messages[-1]]

                result_messages.append(
                    {
                        "role": "assistant",
                        "content": "```json\n"
                        + json.dumps(tool_args, indent=2, ensure_ascii=True)
                        + "\n```",
                        "metadata": {
                            "parent_id": f"tool_call_{tool_name}",
                            "id": f"params_{tool_name}",
                            "title": "Tool Parameters",
                        },
                    }
                )
                partial_messages.append(result_messages[-1])
                yield [result_messages[-1]]

                print(f"Calling tool: {tool_name} with args: {tool_args}")
                try:
                    # Check if session is still valid
                    if not self.session or not self.stdio or not self.write:
                        raise Exception(
                            "MCP session is not connected or has been closed"
                        )

                    result = await self.session.call_tool(tool_name, tool_args)
                except Exception as e:
                    error_msg = f"Error calling tool {tool_name}: {str(e)}"
                    print(error_msg)
                    result_messages.append(
                        {
                            "role": "assistant",
                            "content": f"Sorry, I encountered an error while calling the tool: {error_msg}. Please try again or reload the page.",
                            "metadata": {
                                "title": f"Tool Error for {tool_name.replace('avsolatorio_test_data_mcp_server', '')}",
                                "status": "done",
                                "id": f"error_{tool_name}",
                            },
                        }
                    )
                    partial_messages.append(result_messages[-1])
                    yield [result_messages[-1]]
                    partial_messages = []
                    continue

                if result_messages and "metadata" in result_messages[-2]:
                    result_messages[-2]["metadata"]["status"] = "done"

                result_messages.append(
                    {
                        "role": "assistant",
                        "content": "Here are the results from the tool:",
                        "metadata": {
                            "title": f"Tool Result for {tool_name.replace('avsolatorio_test_data_mcp_server', '')}",
                            "status": "done",
                            "id": f"result_{tool_name}",
                        },
                    }
                )
                partial_messages.append(result_messages[-1])
                yield [result_messages[-1]]
                partial_messages = []

                result_content = result.content
                print(result_content)
                if isinstance(result_content, list):
                    result_content = [r.model_dump() for r in result_content]

                    for r in result_content:
                        # Remove annotations field from each item if it exists
                        r.pop("annotations", None)
                        try:
                            r["text"] = json.loads(r["text"])
                        except:
                            pass

                    print("result_content", result_content)

                result_messages.append(
                    {
                        "role": "assistant",
                        "content": "```\n"
                        + json.dumps(result_content, indent=2)
                        + "\n```",
                        "metadata": {
                            "parent_id": f"result_{tool_name}",
                            "id": f"raw_result_{tool_name}",
                            "title": "Raw Output",
                        },
                    }
                )
                partial_messages.append(result_messages[-1])
                yield [result_messages[-1]]
                partial_messages = []

                claude_messages.append(
                    {"role": "assistant", "content": [content.model_dump()]}
                )
                claude_messages.append(
                    {
                        "role": "user",
                        "content": [
                            {
                                "type": "tool_result",
                                "tool_use_id": tool_id,
                                "content": json.dumps(result_content, indent=2),
                            }
                        ],
                    }
                )

                try:
                    next_response = self.anthropic.messages.create(
                        model=LLM_MODEL,
                        system=SYSTEM_PROMPT,
                        max_tokens=1000,
                        messages=claude_messages,
                        tools=self.tools,
                    )
                    auto_calls += 1
                except OverloadedError:
                    yield [
                        {
                            "role": "assistant",
                            "content": "The LLM API is overloaded now, try again later...",
                        }
                    ]

                print("next_response", next_response.content)

                contents.extend(next_response.content)

    async def _process_query(
        self,
        message: str,
        history: List[Union[Dict[Any, Any], ChatMessage]],
        previous_response_id: str = None,
    ):
        if LLM_PROVIDER == "anthropic":
            async for partial in self._process_query_anthropic(message, history):
                yield partial
        elif LLM_PROVIDER == "openai":
            try:
                async for partial in self._process_query_openai(
                    message, history, previous_response_id
                ):
                    yield partial
            except openai.APIError as e:
                print(e)
                yield [
                    {
                        "role": "assistant",
                        "content": "The LLM encountered an error. Please try again or reload the page.",
                    }
                ]
            except Exception as e:
                print(e)
                yield [
                    {
                        "role": "assistant",
                        "content": f"Sorry, I encountered an unexpected error: `{e}`. Please try again or reload the page.",
                    }
                ]


def gradio_interface(
    server_path_or_url: str = "https://avsolatorio-test-data-mcp-server.hf.space/gradio_api/mcp/sse",
):
    # server_path_or_url = "https://avsolatorio-test-data-mcp-server.hf.space/gradio_api/mcp/sse"
    # server_path_or_url = "wdi_mcp_server.py"

    client = MCPClientWrapper()
    custom_css = """
    .gradio-container {
        background-color: #fff !important;
    }
    .message-row.panel.bot-row {
        background-color: #fff !important;
    }
    .message-row.panel.user-row {
        background-color: #fff !important;
    }
    .user {
        background-color: #f1f6ff !important;
    }
    .bot {
        background-color: #fff !important;
    }
    .role {
        margin-left: 10px !important;
    }
    footer{display:none !important}
    """

    # Disable auto-dark mode by setting theme to None
    with gr.Blocks(title="WDI MCP Client", css=custom_css, theme=None) as demo:
        try:
            gr.Markdown("# Data360 Chat [Prototype]")
            # gr.Markdown("Connect to the WDI MCP server and chat with the assistant")

            with gr.Accordion(
                "Connect to the WDI MCP server and chat with the assistant",
                open=False,
                visible=server_path_or_url.endswith(".py"),
            ):
                with gr.Row(equal_height=True):
                    with gr.Column(scale=4):
                        server_path = gr.Textbox(
                            label="Server Script Path",
                            placeholder="Enter path to server script (e.g., wdi_mcp_server.py)",
                            value=server_path_or_url,
                        )
                    with gr.Column(scale=1):
                        connect_btn = gr.Button("Connect")

                status = gr.Textbox(label="Connection Status", interactive=False)

            chatbot = gr.Chatbot(
                value=[],
                height="81vh",
                type="messages",
                show_copy_button=False,
                avatar_images=("img/small-user.png", "img/small-robot.png"),
                autoscroll=True,
                layout="panel",
                placeholder="Ask development data questions!",
            )
            previous_response_id = gr.State(None)

            with gr.Row(equal_height=True):
                msg = gr.Textbox(
                    label=None,
                    placeholder="Ask about what indicators are available for a specific topic (e.g., What's the definition of GDP?)",
                    scale=4,
                    show_label=False,
                )
                # clear_btn = gr.Button("Clear Chat", scale=1)

            # connect_btn.click(client.connect, inputs=server_path, outputs=status)
            # Automatically call client.connect(...) as soon as the interface loads
            if LLM_PROVIDER == "anthropic":
                demo.load(
                    fn=client.connect,
                    inputs=server_path,
                    outputs=status,
                    show_progress="full",
                )

            msg.submit(
                client.process_message,
                [msg, chatbot, previous_response_id],
                [chatbot, msg, previous_response_id],
                concurrency_limit=10,
            )
            # clear_btn.click(lambda: [], None, chatbot)

        except KeyboardInterrupt:
            if LLM_PROVIDER == "anthropic":
                print("Keyboard interrupt received. Disconnecting from MCP server...")
                asyncio.run(client.disconnect())
            raise KeyboardInterrupt
        # demo.unload(client.disconnect)

    return demo


if __name__ == "__main__":
    if not os.getenv("ANTHROPIC_API_KEY"):
        print(
            "Warning: ANTHROPIC_API_KEY not found in environment. Please set it in your .env file."
        )

    # interface = gradio_interface(server_path_or_url="wdi_mcp_server.py")
    interface = gradio_interface(
        server_path_or_url="https://avsolatorio-test-data-mcp-server.hf.space/gradio_api/mcp/sse"
    )
    interface.launch(
        server_name=os.getenv("SERVER_NAME", "127.0.0.1"),
        server_port=os.getenv("SERVER_PORT", 7860),
        debug=True,
    )