Share a local MCP server
Give Claude, ChatGPT or any MCP client a public HTTPS URL for the MCP server on your laptop.
The Model Context Protocol (MCP) lets AI assistants call tools and read
data from your own servers. While you build an MCP server it runs on
localhost — which cloud-hosted assistants (Claude.ai custom connectors,
ChatGPT, hosted agents) can't reach. mylocalon gives it a public HTTPS URL in
one command, so you can test a remote MCP server without deploying it.
mylocalon http 8000 --subdomain my-mcp Forwarding https://my-mcp.mylocalon.com → http://localhost:8000Your MCP endpoint is now https://my-mcp.mylocalon.com/mcp.
What works
mylocalon forwards plain HTTP requests and delivers each response when it is complete. That matches MCP's Streamable HTTP transport in request/response mode:
| MCP setup | Works through mylocalon? |
|---|---|
Streamable HTTP, JSON responses (application/json) |
✅ Yes — recommended |
| Streamable HTTP, a short SSE response per request that then closes | ✅ Yes — delivered when the response finishes |
Stateless or session-based (Mcp-Session-Id header) |
✅ Yes — headers pass through unchanged |
A long-lived GET event stream for server-initiated messages |
❌ No — it never completes, so it times out after 30 s |
Legacy HTTP+SSE transport (separate /sse endpoint) |
❌ No — needs a streaming connection |
stdio servers (launched as a local process) |
❌ Not over a network — wrap them with an HTTP transport first |
Tool calls must finish within 30 seconds and responses stay under 10 MB (Limits).
Python (official MCP SDK, FastMCP)
# server.py
from mcp.server.fastmcp import FastMCP
mcp = FastMCP("demo", stateless_http=True, json_response=True)
@mcp.tool()
def add(a: int, b: int) -> int:
"""Add two numbers."""
return a + b
if __name__ == "__main__":
mcp.run(transport="streamable-http") # serves http://localhost:8000/mcppython server.py
mylocalon http 8000 --subdomain my-mcpTypeScript (official MCP SDK)
Use the Streamable HTTP transport with JSON responses enabled:
import { StreamableHTTPServerTransport } from "@modelcontextprotocol/sdk/server/streamableHttp.js";
const transport = new StreamableHTTPServerTransport({
sessionIdGenerator: undefined, // stateless
enableJsonResponse: true, // plain JSON instead of an event stream
});Mount it on POST /mcp in your HTTP server (Express, Hono, Node http), then
run mylocalon http <port>.
Connect an AI client
Test it first with the MCP Inspector:
npx @modelcontextprotocol/inspectorChoose Streamable HTTP, enter
https://my-mcp.mylocalon.com/mcp, and list the tools.Add it to your assistant as a remote / custom connector using the same URL — for example in Claude: Settings → Connectors → Add custom connector.
Watch the calls arrive in your terminal as the assistant uses your tools:
10:42:07 POST /mcp 200 38ms
Tip: reserve the subdomain (Reserved subdomains) so the connector URL never changes between sessions.
Secure it
A tunnel URL is public: anyone who has it can call your tools. A hard-to-guess subdomain is not protection.
- Require a secret on every request — check an
Authorization: Bearer …header in your server and reject everything else. - Never expose tools that run shell commands or touch real data without auth.
- Stop the tunnel (
Ctrl+C) when you finish testing.
Troubleshooting
| Symptom | Fix |
|---|---|
| Client hangs, then 504 | The server opened a long-lived event stream. Enable JSON responses (json_response=True / enableJsonResponse: true). |
404 on /mcp |
Check the path your server mounts; FastMCP uses /mcp by default. |
| 502 | Your MCP server isn't running on the port you tunneled. |
| Tools list is empty | Test with the MCP Inspector to separate server bugs from client setup. |