MultiServerMCPClient stdio transport fails to launch installed packages

[amaurygelin]

Description:
I am encountering a Connection closed error when using MultiServerMCPClient to connect to a local MCP server installed in my current virtual environment.

I am running a standard python -m <package> command via the stdio transport. The package (mcp-server-fetch) is installed and verified working in the active virtual environment. However, when MultiServerMCPClient attempts to initialize the connection, it crashes immediately.

Reproduction Steps:

1. Install a simple MCP server package (e.g., mcp-server-fetch) in a virtual environment.
2. Run the following client code from within the same virtual environment:

import asyncio
from langchain_mcp_adapters.client import MultiServerMCPClient

async def main():
    client = MultiServerMCPClient(
        {
            "fetch": {
                "transport": "stdio",
                "command": "python",
                "args": ["-m", "mcp_server_fetch"],
            },
        }
    )

    # This line raises the exception
    tools = await client.get_tools()
    print("Available tools:", tools)

if __name__ == "__main__":
    asyncio.run(main())

Expected Behavior:
The client should successfully start the subprocess using the active environment, import the module mcp_server_fetch, and retrieve the tools.

Actual Behavior:
The execution fails with a Connection closed error during initialization.

Error Log:

  + Exception Group Traceback (most recent call last):
  ...
  |   File "...\client.py", line 17, in main
  |     tools = await client.get_tools()
  |             ^^^^^^^^^^^^^^^^^^^^^^^^
  ...
  |   File "...\mcp\client\session.py", line 153, in initialize
  |     result = await self.send_request(
  |              ^^^^^^^^^^^^^^^^^^^^^^^^
  |   File "...\mcp\shared\session.py", line 288, in send_request
  |     raise McpError(response_or_error.error)
  | mcp.shared.exceptions.McpError: Connection closed

Environment:

• OS: Windows (MINGW64)
• Python: 3.12.10
• Package: mcp-server-fetch (installed in venv)
• Library: langchain-mcp-adapters

Additional Context:

• Running pip show mcp-server-fetch confirms the package is installed in the active .venv.
• This issue seems related to how _create_stdio_session handles environment variables; if it passes env=None (or an empty dict) to the subprocess, the virtual environment activation is lost, and the python command fails to find the installed module.
• Same error when using the command uvx mcp-server-fetch.

[amaurygelin]

Hello! My team and I made some progress on this issue. We found that it fails on a Windows development environment but it works on a WSL environment.
Is this OS issue known? @sydney-runkle

At our company, we’d like to promote the use of langchain-mcp-adapters to enable local MCP servers in our LangChain/LangGraph/Deepagents applications. However, this problem is a major blocker, as most of our data scientists use Windows without WSL.