flat rock
Hi!
I've been experimenting with the Model Context Protocol (MCP) integration and noticed that it currently only supports the legacy SSE-based setup.
Is there support for the newer /mcp HTTP endpoint with bidirectional streaming, or is that still in progress?
Just making sure I didn’t miss something. Thanks!
delicate matrix
me too!... but with no luck
If you want to use that MCP server capability from HA with other tools like Flowise or n8n.. it´s mandatory that you deploy a proper MPC server as described in HA DOC "Install mcp-proxy following the instructions in the README. For example, uv tool install git+https://github.com/sparfenyuk/mcp-proxy. "
"Since most clients do not support native remote servers, you need an additional local MCP server remote gateway."
so I've been trying to deploy that mcp-proxy using portainer in a docker with no luck.. I don't know why I'm having restarts within the image once it´s deployed...
version: "3"
services:
mcp-proxy:
image: ghcr.io/sparfenyuk/mcp-proxy:latest
container_name: mcp-proxy
command: >
--port=8096
--headers Authorization "Bearer YOUR_LONG_TERM_HA_TOKEN"
http://YOUR_HA_IP:8123/mcp_server/sse
ports:
- "8096:8096"
networks:
- default
restart: unless-stopped
networks:
default:
driver: bridge
I've finally got an MCP server up and running using Wrangler and TypeScript. It registers several tools, can save data to a PostgreSQL database, and read from it, making the data accessible via Voice Assist. This is just the beginning, and I have a lot more to implement, with plenty of things that could still go wrong.
delicate matrix
I've finally got an MCP server up and running using Wrangler and TypeScript. It ...
cool! may I ask how did you get it? I would really appreciate any help about this
flat rock
cool! may I ask how did you get it? I would really appreciate any help about thi...
It's still very shaky, at the moment I have broke it by adding more complex features but will probably share everything on github once everything is done.
This is not a complete setup since I have struggle with so many things and diffrent solutions and are not yet sure what to keep. At some point I need to do a fresh reinstall and see all needed steps.
I'm just a newbee in this world.
Create a free Cloudflare account.
On Windows, CMD and navicate to the folder were you want the MCP server.
Execute npm install -g wrangler
You may first need to install npm, nodeJS etc.
Run wrangler login, this will open Cloudflare page to assign your account.
Create a wrangler.toml file with below.
name = "My Secret Project"
main = "scripts/my_tools.ts"
compatibility_date = "2025-07-01"
compatibility_flags = [ "nodejs_compat" ]
[vars]
NODE_ENV = "development"
DATABASE_URL = "postgresql://mcp_user:mcp_password@localhost:5432/mcp_database"
[[migrations]]
new_sqlite_classes = ["MyMCP"]
tag = "v1"
[durable_objects]
bindings = [
{ name = "MCP_OBJECT", class_name = "MyMCP" }
]
[dev]
port = 8790
import { McpAgent } from "agents/mcp";
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { z } from "zod";
export class MyMCP extends McpAgent<Env, Record<string, never>, Record<string, never>> {
server = new McpServer({
name: "My MCP Server",
version: "1.0.0",
});
async init() {
this.server.tool(
"my_mcp_tool",
z.object({}),
async () => {
return {
content: [
{
type: "text",
text: "Hello, this is the response from the tool"
}
]
};
}
);
}
}
export default {
fetch(request: Request, env: Env, ctx: ExecutionContext) {
const url = new URL(request.url);
if (url.pathname === "/sse" || url.pathname === "/sse/message") {
return MyMCP.serveSSE("/sse").fetch(request, env, ctx);
}
if (url.pathname === "/mcp") {
return MyMCP.serve("/mcp").fetch(request, env, ctx);
}
return new Response("Not found", { status: 404 });
},
};
Run set DATABASE_URL=postgresql://mcp_user:mcp_password@localhost:5432/mcp_database before starting the wrangler, ihave it in a bat file.
Start the wrangler by run wrangler dev --ip 10.5.0.10 (Your ip the wrangler is running on)
If you want a sql server.
services:
postgres:
image: postgres:17-alpine
container_name: mcp-sql-server
restart: always
environment:
POSTGRES_USER: mcp_user
POSTGRES_PASSWORD: mcp_password
POSTGRES_DB: mcp_database
ports:
- "5432:5432"
volumes:
- postgres_data:/var/lib/postgresql/data
healthcheck:
test: ["CMD", "pg_isready", "-U", "mcp_user", "-d", "mcp_database"]
interval: 10s
timeout: 5s
retries: 5
volumes:
postgres_data:The database connection.
import postgres from "postgres";
let dbInstance: postgres.Sql | null = null;
export function getDb(databaseUrl: string): postgres.Sql {
if (!dbInstance) {
dbInstance = postgres(databaseUrl, {
max: 10,
idle_timeout: 30,
connect_timeout: 15,
prepare: true,
});
}
return dbInstance;
}
export async function closeDb(): Promise<void> {
if (dbInstance) {
try {
await dbInstance.end();
} catch (error) {
console.error('Error closing database connection:', error);
} finally {
dbInstance = null;
}
}
}
thank you so much for your detailed answer...I really appreciate that.. but it sadly it exceeds my programming skills 😦
I'm trying to solve this using some available tools and things like n8n or flowise.. and even with that it´s been quite complex..
flat rock
At first, I couldn’t find any way to get it working myself, so I’m very happy to finally have made some progress. I ran another test using Python, which feels more flexible for my needs after running into issues with TypeScript.
I created a Python STDIO server and exposed it as SSE with this command:
mcp-proxy --debug --sse-port=8080 python mcp_server.py
You can then point the HA MCP integration at http://<ip>:8080/sse, open the OpenAI integration, and select tooltest-server.
Below is a minimal Python script that replies when the tooltest tool is invoked.
#!/usr/bin/env python3
import asyncio
import json
from mcp.server import Server, NotificationOptions
from mcp.server.models import InitializationOptions
import mcp.types as types
import mcp.server.stdio
server = Server("tooltest-server")
@server.list_tools()
async def list_tools() -> list[types.Tool]:
return [
types.Tool(
name="tooltest",
description="Returns a static test message",
inputSchema={
"type": "object",
"properties": {},
"additionalProperties": False
}
)
]
@server.call_tool()
async def call_tool(name: str, arguments: dict | None) -> list[types.TextContent]:
if name == "tooltest":
return [
types.TextContent(
type="text",
text="Test successful!"
)
]
raise ValueError(f"Unknown tool: {name}")
async def main():
async with mcp.server.stdio.stdio_server() as (read_stream, write_stream):
await server.run(
read_stream,
write_stream,
InitializationOptions(
server_name="tooltest-server",
server_version="1.0.0",
capabilities=server.get_capabilities(
notification_options=NotificationOptions(),
experimental_capabilities={}
),
),
)
if __name__ == "__main__":
asyncio.run(main())
fossil kraken
me too!... but with no luck
If you want to use that MCP server capability from H...
Hi. Trying to do this. Why would one require an MCP Proxy? I have n8n and Home Assistant self hosted on the same network. Right now my issue is that it finds the service at /mcp_server/sse but getting 404: Unathorised. I created a Long Live Token and povided that as the API Key with Key = API_ACCESS_TOKEN. Any insights would be appreciated.