Environment
- agent-framework-core: 1.0.0rc5
- mcp (Python SDK): latest
- MCP server: mongodb-mcp-server@1.9.0 (via npx -y mongodb-mcp-server)
- Python: 3.14
- OS: Windows 11
Description
MCPStdioTool.call_tool hangs indefinitely when the MCP server sends a resources/updated notification alongside the
tool result. The underlying mcp Python SDK (ClientSession.call_tool) completes successfully with the same server and
arguments.
Steps to Reproduce
import asyncio
import os
from agent_framework import MCPStdioTool
from dotenv import load_dotenv
load_dotenv(override=False)
async def main():
env = dict(os.environ)
env["MDB_MCP_CONNECTION_STRING"] = os.getenv("MDB_MCP_CONNECTION_STRING")
mcp = MCPStdioTool(
name="MongoDB",
command="npx",
args=["-y", "mongodb-mcp-server"],
env=env,
load_prompts=False,
request_timeout=30,
)
async with mcp:
print(f"Connected, {len(mcp.functions)} tools available")
tool = {fn.name: fn for fn in mcp.functions}["list-collections"]
# This hangs indefinitely:
result = await tool.invoke(arguments={"database": "sample_mflix"})
print(result)
asyncio.run(main())
Expected Behavior
list-collections returns the result within the request_timeout window.
Actual Behavior
The call hangs indefinitely. Logs show the server responds and sends a resource notification, but the framework never
returns:
INFO:agent_framework._mcp:[server]: Detected a MongoDB connection string in the configuration, trying to connect...
INFO:agent_framework._mcp:[server]: Server with version 1.9.0 started with transport StdioServerTransport
INFO:agent_framework:Function name: list-collections
INFO:agent_framework._mcp:[resources]: Resource updated: debug://mongodb
Workaround
Using the mcp SDK directly (which agent-framework depends on) works without issue:
from mcp.client.session import ClientSession
from mcp.client.stdio import StdioServerParameters, stdio_client
async with stdio_client(server=StdioServerParameters(...)) as transport:
async with ClientSession(read_stream=transport[0], write_stream=transport[1]) as session:
await session.initialize()
result = await session.call_tool("list-collections", {"database": "sample_mflix"})
# Returns immediately with correct result
Analysis
The mongodb-mcp-server sends a resources/updated notification (debug://mongodb) during tool execution. The framework's
call_tool method (_mcp.py:1031) or the result parsing pipeline appears to not complete when these interleaved
notifications arrive, even though the underlying ClientSession handles them correctly.
Environment
Description
MCPStdioTool.call_tool hangs indefinitely when the MCP server sends a resources/updated notification alongside the
tool result. The underlying mcp Python SDK (ClientSession.call_tool) completes successfully with the same server and
arguments.
Steps to Reproduce
import asyncio
import os
from agent_framework import MCPStdioTool
from dotenv import load_dotenv
load_dotenv(override=False)
async def main():
env = dict(os.environ)
env["MDB_MCP_CONNECTION_STRING"] = os.getenv("MDB_MCP_CONNECTION_STRING")
asyncio.run(main())
Expected Behavior
list-collections returns the result within the request_timeout window.
Actual Behavior
The call hangs indefinitely. Logs show the server responds and sends a resource notification, but the framework never
returns:
INFO:agent_framework._mcp:[server]: Detected a MongoDB connection string in the configuration, trying to connect...
INFO:agent_framework._mcp:[server]: Server with version 1.9.0 started with transport StdioServerTransport
INFO:agent_framework:Function name: list-collections
INFO:agent_framework._mcp:[resources]: Resource updated: debug://mongodb
Workaround
Using the mcp SDK directly (which agent-framework depends on) works without issue:
from mcp.client.session import ClientSession
from mcp.client.stdio import StdioServerParameters, stdio_client
async with stdio_client(server=StdioServerParameters(...)) as transport:
async with ClientSession(read_stream=transport[0], write_stream=transport[1]) as session:
await session.initialize()
result = await session.call_tool("list-collections", {"database": "sample_mflix"})
# Returns immediately with correct result
Analysis
The mongodb-mcp-server sends a resources/updated notification (debug://mongodb) during tool execution. The framework's
call_tool method (_mcp.py:1031) or the result parsing pipeline appears to not complete when these interleaved
notifications arrive, even though the underlying ClientSession handles them correctly.