Server is Shutdown when hot reload (FastMCP + FastAPI)

[chungtran4078]

Description

I added FastAPI into my code to perform reload when something changed.
My server.py:

from fastmcp import FastMCP
from fastapi import FastAPI


app = FastAPI()
mcp = FastMCP.from_fastapi(app, "APCTools")


@mcp.resource(uri="apc://policy", name="The Policy of APC")
async def read_policy() -> str:
    return json.dumps({
        "status": "success",
        "policy": "The policy of APC company allows employees to sleep freely during working hours, receive a salary increase once a month, and arrive late up to 40 hours per week.",
        "error": ""
    })

app.mount("/", mcp.sse_app())

Start server command: uvicorn app.server:app --host 0.0.0.0 --port 8000 --reload

Everything fine till I perform save file server.py, this time server will reload and shutdown

postgres-mcp-server-container  | INFO:     Waiting for application startup.
postgres-mcp-server-container  | INFO:     Application startup complete.
postgres-mcp-server-container  | INFO:     192.168.65.1:41967 - "GET /sse HTTP/1.1" 200 OK
postgres-mcp-server-container  | INFO:     192.168.65.1:26200 - "POST /messages/?session_id=a44e28b433964125b93e57f22692fe6a HTTP/1.1" 202 Accepted
postgres-mcp-server-container  | INFO:     192.168.65.1:26200 - "POST /messages/?session_id=a44e28b433964125b93e57f22692fe6a HTTP/1.1" 202 Accepted
postgres-mcp-server-container  | INFO:     192.168.65.1:26200 - "POST /messages/?session_id=a44e28b433964125b93e57f22692fe6a HTTP/1.1" 202 Accepted
postgres-mcp-server-container  | 2025-05-02 01:55:45,368 - INFO - Processing request of type ListResourcesRequest
postgres-mcp-server-container  | WARNING:  StatReload detected changes in 'server.py'. Reloading...
postgres-mcp-server-container  | INFO:     Shutting down
postgres-mcp-server-container  | INFO:     Waiting for connections to close. (CTRL+C to force quit)

Example Code

Version Information

FastMCP version: 2.2.6
MCP version: 1.6.0
Python version: 3.13.3
Platform: macOS-15.1-x86_64-i386-64bit-Mach-O

Additional Context

No response

[jlowin]

The low-level MCP server does not have the ability to shut down gracefully, and a new server is started for every SSE connection - that is why it is hanging. I have a branch in flight that will solve this for a different purpose.

[jlowin]

Actually this is more challenging to fix without modifying the low level server - however unfortunately it is not a FastMCP issue, as much as I would like to solve it as well.

[westonplatter]

I am seeing the same issue + discussion in the mcp python repo, modelcontextprotocol/python-sdk#514. This pull request was one of the proposed solutions, modelcontextprotocol/python-sdk#586. Cross posting in case it's helpful for others.

[chungtran4078]

Thanks @jlowin @westonplatter a lot. I think i will wait for next release

[westonplatter]

Ah ha! Found a work around solution per @mattmess1221's comment. I was able to change the file, see the fastapi mcp app reboot, and reconnect to it using MCP Inspector v0.11.0.

# where my fastapi app is located in app/main.py as app
UVICORN_TIMEOUT_GRACEFUL_SHUTDOWN=0 uvicorn app.main:app --reload 

Using UVICORN_TIMEOUT_GRACEFUL_SHUTDOWN with the fastapi command did not work for me.

[jlowin]

@westonplatter while we wait for low-level improvements, #323 allows you to pass a graceful timeout shutdown param directly without the env var, and defaults it to 0. Hopefully that helps!

[leavor]

downgrade uvicorn

pip install uvicorn==0.21.1