NotImplementedError with MultiServerMCPClient on Windows due to SelectorEventLoop

[woaidapaopao]

重现步骤 (To Reproduce)

1. 在 Windows 11 上，使用 Python 3.11。
2. 配置 MultiServerMCPClient 以通过 stdio 启动子进程，例如：

   @asynccontextmanager
    async def make_graph():
        async with MultiServerMCPClient(
            {
                "math": {
                    "command": "D:\codehub\langgraph_test\.venv\Scripts\python.exe",
                    # Make sure to update to the full absolute path to your math_server.py file
                    "args": ["D:\codehub\langgraph_test\mcp\math_server.py"],
                    "transport": "stdio",
                },
                "weather": {
                    # make sure you start your weather server on port 8000
                    "url": "http://localhost:8000/sse",
                    "transport": "sse",
                }
            }
        ) as client:
            agent = create_react_agent(model, client.get_tools())
            yield agent

3. 使用langgraph dev运行代码。
4. 观察到以下错误：

   File "D:\Python\Python311\Lib\asyncio\base_events.py", line 503, in _make_subprocess_transport
     raise NotImplementedError
   NotImplementedError
   

环境信息 (Environment)

• 操作系统: Windows 11
• Python 版本: 3.11
• 依赖版本:
  • langchain-mcp-adapters: 0.0.4

错误堆栈

Traceback (most recent call last):
File "D:\codehub\langgraph_test.venv\Lib\site-packages\langchain_mcp_adapters\client.py", line 223, in aenter
await self.connect_to_server_via_stdio(server_name, **connection_dict)
File "D:\codehub\langgraph_test.venv\Lib\site-packages\langchain_mcp_adapters\client.py", line 176, in connect_to_server_via_stdio
stdio_transport = await self.exit_stack.enter_async_context(stdio_client(server_params))
File "D:\Python\Python311\Lib\asyncio\base_events.py", line 503, in _make_subprocess_transport
raise NotImplementedError
NotImplementedError

[jayanthma]

Hi @woaidapaopao, Have you sorted this issue?

[ibbybuilds]

Any update on this? @jayanthma @woaidapaopao @hanfi @eyurtsev @LicsDaSheng

[chenmohan123]

NotImplementedError in asyncio subprocess_exec when using MCP client on Windows with Python 3.12
stdio_client not starting in windows: getting NotImplementedError
官方项目也有相关错误。大致是由于 asyncio 在 Windows 系统上尝试创建子进程时触发了 NotImplementedError。这通常是因为在某些情况下，asyncio 的子进程支持在 Windows 上可能会受到限制或未完全实现。

[jayanthma]

Hi @mibrahim2001, the issue was persisting in the jupyter notebook so I converted it to .py and it ran successfully. Find the code below.

from langchain_mcp_adapters.client import MultiServerMCPClient

from langgraph.prebuilt import create_react_agent

from langchain_aws import ChatBedrockConverse
import boto3

bedrock_client=boto3.client(service_name='bedrock-runtime', region_name="us-east-1")

model = ChatBedrockConverse(
client=bedrock_client,
model="us.meta.llama3-3-70b-instruct-v1:0",
temperature=0,
)

async def test_mcp_client():
async with MultiServerMCPClient(
{
"math": {
"command": "C:/Users/jay/coding_projects/LangGraph/venv/Scripts/python.exe",
# Make sure to update to the full absolute path to your math_server.py file
"args": ["C:/Users/jay/coding_projects/LangGraph/mcp/math_server.py"],
"transport": "stdio",
},
"weather": {
# make sure you start your weather server on port 8000
"url": "http://localhost:8000/sse",
"transport": "sse",
}
}
) as client:
tools = client.get_tools()
print("Tools: ", tools)
agent = create_react_agent(model, client.get_tools(), debug=True)
math_response = await agent.ainvoke({"messages": "what's 3*175"})
weather_response = await agent.ainvoke({"messages": "what is the weather in nyc?"})

return math_response, weather_response

if name == "main":
import asyncio
math_response, weather_response = asyncio.run(test_mcp_client())
print(math_response)
print(weather_response)

Hope it helps!!

[VinyVan]

Hi , @woaidapaopao.
It seems like you are using langgraph CLI , so to fix the issue , you must just add the '--no-reload' arg .
When dev mode , the command must be 'langgraph dev --no-reload ' or 'langgraph dev --no-reload --allow-blocking ' to allow blocking.

the issue is due to the reloading behavior.
For MCP transport type 'sse' there is no issue , the issue occured only for 'stdio' .

the above commands resolve the problem .

[xpilasneo4j]

Using exactly the code from the README.md, I have the same error on Windows.
Seems there are some incompatibilities between anyio, asyncio and Windows: if we try to force the loop to be WindowsProactorEvent loop, it doesn't work as there is another issue with fileno...
I will try the code suggested by @jayanthma
Maybe it would be good to have the code in the README tested on all majors platforms?
If it is not the case or if Windows won't be supported, mention it... that people knows
Thank you for all the work on this amazing lib!

[LastHopeOfGPNU]

This works for me:

import asyncio
asyncio.set_event_loop_policy(asyncio.WindowsProactorEventLoopPolicy())

[xpilasneo4j]

It worked because the error moved to another part of the code: I got an error on the fileno after

[brisacoder]

I had similar problems running within Jupyter Notebooks, running natively I have no problems.

I have Python 3.13 and Windows 11. You do not need to force the set_event_loop_policy() in later versions of Python since the platform is detected automatically. Here is how I connect to MCP Servers and load the tools. I tested from within a node in the graph and outside as I create a react_agent() or Tool node.

async def load_servers(config_path: str) -> Dict:
    """Load server configuration from a JSON file (typically mcp_config.json)
    and creates an instance of each server (no active connection until 'start' though).

    Args:
        config_path: Path to the JSON configuration file.
    """
    async with aiofiles.open(config_path, "r", encoding="utf-8") as config_file:
        content = await config_file.read()
        config = json.loads(content)

    return config


async def get_mcp_tools() -> List[StructuredTool]:
    """Main function to run the MCP client."""
    mcp_servers = await load_servers("mcp_config.json")

    client = MultiServerMCPClient(mcp_servers["mcpServers"])
    all_tools = await client.get_tools()
    return all_tools

[theshewaspretty]

I tested this case in a windows environment.
using python 3.11.0
langchain-mcp-adapters==0.0.9

It works fine when I specify sys.platform as win32.

Detecting if we're running on Windows using sys.platform == 'win32'
Setting the appropriate event loop policy using asyncio.set_event_loop_policy(asyncio.WindowsProactorEventLoopPolicy())

import sys

if sys.platform == 'win32':
    asyncio.set_event_loop_policy(asyncio.WindowsProactorEventLoopPolicy())

****


import logging

logging.basicConfig(
    level=logging.INFO,  # Default to INFO level
    format='%(filename)s:%(lineno)d | %(message)s',
    handlers=[
        logging.StreamHandler(sys.stderr)
    ]
)

[eyurtsev]

Should have been resolved in recent versions of the MCP sdk modelcontextprotocol/python-sdk#1044