MCP Gateway Connection Drops After Some Time (503+ ClosedResourceError) -Best practices ?

[user1500177]

Hi Community,

I’m using Agno (v2.5.9)+ MCP (via an MCP gateway) and wanted some guidance on connection handling and resilience.

What I did

Initially connected directly to an MCP server → worked fine

After redeploys, I had to manually restart my agent (tool calls failed)

Later had to switch to MCP gateway (via streaming HTTP)

Current Problem

Initially everything works (LLM + tool calls succeed)

After some time, connection drops and I start seeing errors like below

Errors (sanitized, same order)

Error in post_writer
Traceback (most recent call last):
File ".../mcp/client/streamable_http.py", line 571, in post_writer
await handle_request_async()
File ".../mcp/client/streamable_http.py", line 565, in handle_request_async
await self._handle_post_request(ctx)
File ".../mcp/client/streamable_http.py", line 358, in _handle_post_request
response.raise_for_status()
File ".../httpx/_models.py", line 829, in raise_for_status
raise HTTPStatusError(message, request=request, response=self)
httpx.HTTPStatusError: Server error '503 Service Unavailable' for url '/mcp'

Error parsing SSE message
Traceback (most recent call last):
File ".../mcp/client/streamable_http.py", line 229, in _handle_sse_event
await read_stream_writer.send(session_message)
File ".../anyio/streams/memory.py", line 249, in send
self.send_nowait(item)
File ".../anyio/streams/memory.py", line 218, in send_nowait
raise ClosedResourceError
anyio.ClosedResourceError

Error while sending ping / tool call
Traceback (most recent call last):
File ".../agno/utils/mcp.py", line 75, in call_tool
await active_session.send_ping()
File ".../mcp/client/session.py", line 222, in send_ping
return await self.send_request(...)
File ".../mcp/shared/session.py", line 281, in send_request
await self._write_stream.send(...)
File ".../anyio/streams/memory.py", line 249, in send
self.send_nowait(item)
File ".../anyio/streams/memory.py", line 218, in send_nowait
raise ClosedResourceError
anyio.ClosedResourceError

ERROR Traceback (most recent call last):
File ".../agno/utils/mcp.py", line 75, in call_tool
await active_session.send_ping()
File ".../mcp/client/session.py", line 222, in send_ping
return await self.send_request(...)
File ".../mcp/shared/session.py", line 281, in send_request
await self._write_stream.send(...)
File ".../anyio/streams/memory.py", line 249, in send
self.send_nowait(item)
File ".../anyio/streams/memory.py", line 218, in send_nowait
raise ClosedResourceError
anyio.ClosedResourceError

ERROR Failed to call MCP tool '':
Traceback (most recent call last):
File ".../agno/utils/mcp.py", line 80, in call_tool
result = await active_session.call_tool(tool_name, kwargs)
File ".../mcp/client/session.py", line 383, in call_tool
result = await self.send_request(...)
File ".../mcp/shared/session.py", line 281, in send_request
await self._write_stream.send(...)
File ".../anyio/streams/memory.py", line 249, in send
self.send_nowait(item)
File ".../anyio/streams/memory.py", line 218, in send_nowait
raise ClosedResourceError
anyio.ClosedResourceError

I saw that there is a flag called refresh_connection , which can be set while creating the tool instance of MCPTools.

Questions

1. Is it expected to see connection closed errors when using refresh_connection parameter ?

2. Should MCP/tool connections be created at startup or handled dynamically?

3. Is refresh/revalidation the standard approach for handling MCP restarts/outages?

4. Do these errors (503, ClosedResourceError, SSE issues) indicate gateway timeouts/instability? (Or is it any mistake from my part How i connect to the mcp or something ?)

   server_params = StreamableHTTPClientParams(
   url=mcp_gateway_url,
   headers=headers,
   timeout=30.0,
   )
   mcp_gateway_tools = MCPTools(
   server_params=server_params,
   transport="streamable-http",
   timeout_seconds=30,
   refresh_connection=True,
   )
   Later in the lifespan of the FastAPI app i connect to this with mcp_gateway_tools.connect(), and later call the await tool.close() also

5. Any idea about performance impact of frequent connection refresh?