Multiple SSE urls on same port

[fengerzh]

Enhancement Description

I would like to start a FastMCP server on port 8000, but isolate tools into different groups, so for example, if MCP client set url to https://www.example.com:8000/sse1 see tools A, but won't see tools B. And also, if MCP client set url to https://www.example.com:8000/sse2 see tools B and not see tools A. Is that possible?

Use Case

This is very useful as we don't want to start multiple processes and use multiple ports.

Proposed Implementation

[shenqing-Fer]

I used this method to achieve this goal. Hope this can help you. In addition, I also tried mounting the sse_app of different mcp instances to different URLs, for example: Mount('/news', app=mcp1.sse_app()). However, the resulting message_path obtained from /news/sse does not include the /news prefix, which causes the client to fail in accessing the correct URL.

我使用了以下方案达到这个目标，希望对你有所帮助。此外，我也尝试过将不同的mcp的sse_app mount到不同的url上，比如Mount('/news', app=mcp1.sse_app())，但是通过/news/sse获取到的message_path不会加上这个前缀，会导致client无法访问到正确的url。

import uvicorn
from mcp.server.fastmcp import FastMCP
from starlette.applications import Starlette
from starlette.routing import Mount

mcp1 = FastMCP(name="news", sse_path='/news/sse', message_path='/news/messages/')
mcp2 = FastMCP(name="stock", sse_path='/stock/sse', message_path='/stock/messages/')


@mcp1.tool()
async def get_news():
    """Search for today's news.
    """
    return 'James was elected president.'

@mcp2.tool()
async def get_stock():
    """Search for today's stock prices.
    """
    return '$100'

mcp_map = {'news': mcp1.sse_app(),
           'stock': mcp2.sse_app()}

class PrefixMappingMiddleware:
    def __init__(self):
        pass

    async def __call__(self, scope, receive, send):
        path = scope.get("path")
        path_items = path.split('/')
        tag = path_items[1]
        await mcp_map[tag](scope, receive, send)


if __name__ == '__main__':
    app = Starlette(
        routes=[
            Mount('/', app=PrefixMappingMiddleware())
        ]
    )
    uvicorn.run(app, host="0.0.0.0", port=80)

[jlowin]

This is a valid use case but there is currently a bug preventing this in the low-level SDK (see e.g. modelcontextprotocol/python-sdk#386) but once resolved it should be straightforward to mount multiple MCP servers under a single Starlette server as @shenqing-Fer described!