Explore MCP

import asyncio import os import json if "OPENAI_API_KEY" not in os.environ: from utils.helpers import get_api_key os.environ["OPENAI_API_KEY"] = get_api_key() from fastmcp import Client from openai import AsyncOpenAI # --- 1. Client Setup --- LLM_PROVIDER_CLIENT = AsyncOpenAI() # Or use Anthropic, Gemini, local Ollama client, etc. MCP_SERVER_URL = "http://127.0.0.1:8000/mcp" async def orchestrate_llm_agent(user_prompt: str): async with Client(MCP_SERVER_URL) as mcp_client: # --- 2. Tool Discovery & Translation --- mcp_tools = await mcp_client.list_tools() # Convert MCP Tool schema (JSON Schema) to OpenAI's Function Calling format llm_tools = [] for tool in mcp_tools: llm_tools.append({ "type": "function", "function": { "name": tool.name, "description": tool.description, "parameters": tool.inputSchema, # FastMCP provides the JSON Schema } }) messages = [{"role": "user", "content": user_prompt}] # --- 3. Orchestration Loop --- loop_count = 0 while True: # 1. Send query and tools to the LLM response = await LLM_PROVIDER_CLIENT.chat.completions.create( model="gpt-5", messages=messages, tools=llm_tools ) tool_calls = response.choices[0].message.tool_calls if not tool_calls: # LLM decided it has the final answer (or cannot use tools) if loop_count == 0: print(f"No tools found for the prompt: ------- {user_prompt}") return response.choices[0].message.content # 2. LLM requests a tool call: Execute the tool using the FastMCP client messages.append(response.choices[0].message) # Add tool call request to history for call in tool_calls: tool_name = call.function.name arguments = json.loads(call.function.arguments) print(f"Agent called tool: {tool_name} with arguments: {arguments}") # Execute the tool via the FastMCP client! mcp_result = await mcp_client.call_tool(tool_name, arguments) # Get the raw data from the result object tool_output = mcp_result.data # 3. Send the tool result back to the LLM new_tool_msg = { "role": "tool", "tool_call_id": call.id, "name": tool_name, "content": str(tool_output), # Send the output as a string for the LLM to read } messages.append(new_tool_msg) loop_count += 1 if __name__ == '__main__': # Example run print(asyncio.run(orchestrate_llm_agent("Add two numbers 23 and 45"))) print(asyncio.run(orchestrate_llm_agent("subtract two numbers 23 and 45"))) print(asyncio.run(orchestrate_llm_agent("Divide number 23 with number 45"))) print(asyncio.run(orchestrate_llm_agent("Who is the CM of Karnataka, India?")))