MCP sHTTP server (+Cursor integration)

4 months ago 2

TL;DR: I’m in the midst of migrating ToolKami’s MCP server from SSE to sHTTP. Here is a simple setup that you can use with Cursor too!

Below video shows how easy it is to start the MCP server (in a single command) and have Cursor agent use its tool.

The MCP Server

The main server is only <50 lines combined, there are a few interesting things happening here:

You can effectively run this program like a binary with all dependencies through UV script
You can selectively decide which tool makes sense for your current application as they are mounted on different paths Mount("/filesystem/"), ...)
Your agent can self-edit its tool in realtime because uvicorn’s reload=True and reload_dirs=["."] settings hot-reload modules whenever there is a change

# /// script # requires-python = ">=3.10" # dependencies = [ # "requests>=2,<3", # "mcp>=1.2.0,<2", # ] # /// from contextlib import AsyncExitStack, asynccontextmanager import uvicorn from starlette.applications import Starlette from starlette.routing import Mount from shttp_modules.filesystem import mcp as filesystem_mcp def combine_lifespans(*lifespans): @asynccontextmanager async def combined_lifespan(app): async with AsyncExitStack() as stack: for l in lifespans: ctx = l(app) await stack.enter_async_context(ctx) yield return combined_lifespan main_app = Starlette( routes=[ Mount("/filesystem/", app=filesystem_mcp.streamable_http_app()), ], lifespan=combine_lifespans( lambda _: filesystem_mcp.session_manager.run(), ), ) if __name__ == "__main__": uvicorn.run( "__main__:main_app", host="127.0.0.1", port=8002, reload=True, reload_dirs=["."] )

Cursor Client

For Cursor to connect to the server.

Create a new file .cursor/mcp.json

{ "mcpServers": { "toolkami-fs": { "url": "http://127.0.0.1:8002/filesystem/mcp" } } }

Then go to Cursor > Settings > Cursor Settings > Tools & Integrations to enable the tool:

Then your agent will be able to start using your tool like in the first video!

Summary

An implementation of everything discussed above can be found here.

It is pretty easy to provide custom, sharp tools to your agent, the tricky part is deciding which tools they should have access to. For code editing, I learnt that having the right diff_fenced_edit_file is critical to defend against catastrophic forgetting - where AI agents overwrite the entire file with cat and delete perfectly functioning code.

I am currently migrating ToolKami which I am using daily to a new setup and will likely document my findings and turn this into a series. Follow this blog or me on X if you’re interested. Thanks for reading and have a nice day!

Read Entire Article