agent-mcp 0.1.2__py3-none-any.whl → 0.1.4__py3-none-any.whl

agent_mcp/langgraph_mcp_adapter.py

@@ -0,0 +1,325 @@
+"""
+LangGraph MCP Adapter - Adapt LangGraph agents to work with MCP.
+
+This module provides an adapter that allows LangGraph agents to work within
+the Model Context Protocol (MCP) framework, enabling them to collaborate
+with agents from other frameworks like Autogen and CrewAI.
+
+Supports both workflow-based and tool-based LangGraph agents:
+1. Workflow-based: Uses StateGraph for defining agent behavior
+2. Tool-based: Uses LangChain tools and agent executors
+"""
+
+import asyncio
+import traceback
+from typing import Dict, Any, Optional, Callable, List, Union
+from langchain.tools import BaseTool
+from langchain.agents import AgentExecutor
+from langchain_openai import ChatOpenAI
+from langchain.agents import create_openai_tools_agent
+from langchain.prompts import ChatPromptTemplate, MessagesPlaceholder
+from langgraph.graph import StateGraph
+from .mcp_agent import MCPAgent
+from .mcp_transport import HTTPTransport
+from fastapi import FastAPI, Request
+import uvicorn
+from threading import Thread
+import time
+
+class LangGraphMCPAdapter(MCPAgent):
+    """
+    Adapter for LangGraph agents to work with MCP.
+    
+    This adapter supports both:
+    1. Workflow-based agents using StateGraph
+    2. Tool-based agents using LangChain tools
+    """
+    
+    def __init__(
+        self,
+        name: str,
+        workflow: Optional[StateGraph] = None,
+        tools: Optional[List[BaseTool]] = None,
+        process_message: Optional[Callable] = None,
+        transport: Optional[HTTPTransport] = None,
+        client_mode: bool = True,
+        state_type: Optional[type] = None,
+        **kwargs
+    ):
+        """
+        Initialize the LangGraph MCP adapter.
+        
+        Args:
+            name: Name of the agent
+            workflow: Optional StateGraph workflow for workflow-based agents
+            tools: Optional list of tools for tool-based agents
+            process_message: Optional custom message processing function
+            transport: Optional transport layer
+            client_mode: Whether to run in client mode
+            **kwargs: Additional arguments to pass to MCPAgent
+        """
+        # Initialize MCPAgent with transport
+        super().__init__(name=name, transport=transport, **kwargs)
+        
+        if workflow and tools:
+            raise ValueError("Cannot specify both workflow and tools. Choose one pattern.")
+            
+        if workflow:
+            # Workflow-based agent
+            self.workflow = workflow
+            self.state_type = state_type
+            self.executor = None
+        elif tools:
+            # Tool-based agent
+            llm = ChatOpenAI(temperature=0)
+            prompt = ChatPromptTemplate.from_messages([
+                ("system", "You are a helpful AI assistant that can use tools to accomplish tasks."),
+                ("human", "{input}"),
+                MessagesPlaceholder(variable_name="agent_scratchpad")
+            ])
+            agent = create_openai_tools_agent(llm, tools, prompt)
+            self.executor = AgentExecutor.from_agent_and_tools(
+                agent=agent,
+                tools=tools,
+                handle_parsing_errors=True
+            )
+            self.workflow = None
+            self.state_type = None
+        else:
+            raise ValueError("Must specify either workflow or tools")
+        
+        self.custom_process_message = process_message
+        self.transport = transport
+        self.client_mode = client_mode
+        self.task_queue = asyncio.Queue()
+        self.state: Dict[str, Any] = {}
+        self.server_ready = asyncio.Event()
+        
+        # Create FastAPI app for server mode
+        self.app = FastAPI()
+        
+        @self.app.post("/message")
+        async def handle_message(request: Request):
+            return await self._handle_message(request)
+            
+        @self.app.on_event("startup")
+        async def startup_event():
+            self.server_ready.set()
+            
+        self.server_thread = None
+        
+    async def _handle_message(self, request: Request):
+        """Handle incoming HTTP messages"""
+        try:
+            message = await request.json()
+            await self.task_queue.put(message)
+            return {"status": "ok"}
+        except Exception as e:
+            return {"status": "error", "message": str(e)}
+            
+    async def process_messages(self):
+        """Process incoming messages from the transport layer"""
+        while True:
+            try:
+                message, message_id = await self.transport.receive_message()
+                print(f"{self.name}: Received message {message_id}: {message}")
+                
+                if message and isinstance(message, dict):
+                    # Add message_id to message for tracking
+                    message['message_id'] = message_id
+                    
+                    # Standardize message structure
+                    if 'content' not in message and message.get('type') == 'task':
+                        message = {
+                            'type': 'task',
+                            'content': {
+                                'task_id': message.get('task_id'),
+                                'description': message.get('description'),
+                                'type': 'task'
+                            },
+                            'message_id': message_id,
+                            'from': message.get('from', 'unknown')
+                        }
+                    
+                    # --- Idempotency Check ---
+                    if not super()._should_process_message(message):
+                        if message_id and self.transport:
+                            asyncio.create_task(self.transport.acknowledge_message(self.name, message_id))
+                            print(f"[{self.name}] Acknowledged duplicate task {message.get('task_id')} (msg_id: {message_id})")
+                        continue
+                    
+                    if message.get('type') == 'task':
+                        print(f"{self.name}: Queueing task with message_id {message_id}")
+                        await self.task_queue.put(message)
+                    elif self.custom_process_message:
+                        await self.custom_process_message(self, message)
+                    else:
+                        print(f"{self.name}: Unknown message type: {message.get('type')}")
+                        if message_id and self.transport:
+                            await self.transport.acknowledge_message(self.name, message_id)
+                            print(f"{self.name}: Acknowledged unknown message {message_id}")
+            except asyncio.CancelledError:
+                print(f"{self.name}: Message processor cancelled")
+                break
+            except Exception as e:
+                print(f"{self.name}: Error processing message: {e}")
+                traceback.print_exc()
+                await asyncio.sleep(1)
+                
+    async def execute_task(self, task: Dict[str, Any]):
+        """Execute a task using either workflow or executor"""
+        try:
+            if self.workflow:
+                # Always initialize state as a dictionary for LangGraph workflows here.
+                # LangGraph itself handles the state type defined in StateGraph().
+                state_dict = {"message": task, "result": None}
+                
+                # Run workflow
+                try:
+                    print(f"{self.name}: Running workflow with initial state_dict: {state_dict}")
+                    # Compile the workflow if not already compiled
+                    if not hasattr(self, '_compiled_workflow'):
+                        self._compiled_workflow = self.workflow.compile()
+                    # Pass the initial state dict directly
+                    final_state = await self._compiled_workflow.ainvoke(state_dict)
+                    # Use the workflow's final state
+                    result = final_state
+                    print(f"{self.name}: Workflow finished with final_state: {result}") 
+                    return {"result": result, "error": None}
+                except Exception as e:
+                    print(f"Error running workflow: {str(e)}")
+                    import traceback
+                    traceback.print_exc() 
+                    return {"result": None, "error": str(e)}
+            elif self.executor:
+                # Run with executor
+                result = await self.executor.arun(task)
+                return {"result": result, "error": None}
+            else:
+                return {"result": None, "error": "No workflow or executor configured"}
+        except Exception as e:
+            return {
+                "result": f"[FROM EXECUTE_TASK] Error executing task: {str(e)}",
+                "error": True
+            }
+            
+    async def process_tasks(self):
+        """Process tasks from the queue"""
+        while True:
+            try:
+                task = await self.task_queue.get()
+                
+                # Extract task details
+                task_content = task.get('content', task.get('task', {}))
+                task_id = task.get('task_id') or task_content.get('task_id')
+                task_description = task.get('description') or task_content.get('description')
+                message_id = task.get('message_id')
+                reply_to = task.get('reply_to')
+                
+                if not task_id or not task_description:
+                    print(f"[ERROR] {self.name}: Task missing required fields: {task}")
+                    self.task_queue.task_done()
+                    continue
+                
+                print(f"\n{self.name}: Processing task {task_id} with message_id {message_id}")
+                
+                try:
+                    # Execute the task
+                    result = await self.execute_task(task_content)
+                    
+                    # Mark task completed
+                    super()._mark_task_completed(task_id)
+                    
+                    # Send result back if reply_to is specified
+                    if reply_to:
+                        print(f"{self.name}: Sending result back to {reply_to}")
+                        await self.transport.send_message(
+                            reply_to,
+                            {
+                                "type": "task_result",
+                                "task_id": task_id,
+                                "result": result['result'],
+                                "sender": self.name,
+                                "original_message_id": message_id,
+                                "error": result['error']
+                            }
+                        )
+                        print(f"{self.name}: Result sent successfully")
+                        
+                        # Acknowledge task completion
+                        if message_id:
+                            await self.transport.acknowledge_message(self.name, message_id)
+                            print(f"{self.name}: Task {task_id} acknowledged with message_id {message_id}")
+                        else:
+                            print(f"{self.name}: No message_id for task {task_id}, cannot acknowledge")
+                except Exception as e:
+                    print(f"{self.name}: Error processing task: {e}")
+                    traceback.print_exc()
+                    
+                    if reply_to:
+                        await self.transport.send_message(
+                            reply_to,
+                            {
+                                "type": "task_result",
+                                "task_id": task_id,
+                                "result": f"Error: {str(e)}",
+                                "sender": self.name,
+                                "original_message_id": message_id,
+                                "error": True
+                            }
+                        )
+                
+                self.task_queue.task_done()
+                
+            except Exception as e:
+                print(f"{self.name}: Error processing task: {e}")
+                traceback.print_exc()
+                await asyncio.sleep(1)
+                
+    def run(self):
+        """Start the message and task processors"""
+        if not self.transport:
+            raise ValueError(f"{self.name}: No transport configured")
+            
+        # Start the transport server if not in client mode
+        if not self.client_mode:
+            def run_server():
+                config = uvicorn.Config(
+                    app=self.app,
+                    host=self.transport.host,
+                    port=self.transport.port,
+                    log_level="info"
+                )
+                server = uvicorn.Server(config)
+                server.run()
+                
+            self.server_thread = Thread(target=run_server, daemon=True)
+            self.server_thread.start()
+        else:
+            # In client mode, we're ready immediately
+            self.server_ready.set()
+            
+        print(f"{self.name}: Starting message processor...")
+        asyncio.create_task(self.process_messages())
+        
+        print(f"{self.name}: Starting task processor...")
+        asyncio.create_task(self.process_tasks())
+        
+    async def connect_to_server(self, server_url: str):
+        """Connect to a coordinator server"""
+        if not self.client_mode:
+            raise ValueError("Agent not configured for client mode")
+            
+        # Wait for server to be ready before connecting
+        if not self.server_ready.is_set():
+            await asyncio.wait_for(self.server_ready.wait(), timeout=10)
+            
+        # Register with the coordinator
+        await self.transport.send_message(
+            server_url,
+            {
+                "type": "register",
+                "agent_name": self.name,
+                "agent_url": self.transport.get_url()
+            }
+        )