Host a remote MCP server
A local MCP server works for one person on one machine. Hosting it remotely means every client — your laptop, your teammates, an agent running somewhere else — can use the same tools over HTTPS.
What you need
- An MCP server that speaks an HTTP transport (most frameworks support streamable HTTP).
- A server to run it on, with HTTPS — that's the part below.
- Authentication. An MCP server on the open internet with no auth is a public API for your tools.
Deploy it
The agent creates the server, installs your dependencies, starts the MCP server from /root/run.sh so it
restarts on crash, and gives you an HTTPS address like https://my-tools.thevibehosting.com.
Connect a client
Add the URL as a remote MCP server in your client's settings and pass the token as a header. The exact place differs between clients; your agent can do this for you too.
Cold starts: sleep or always-on
A server sleeps after about 10 minutes without incoming traffic. The first request after that wakes it: it gets a
503 with Retry-After: 1 while the server starts, and the retry goes through a few seconds later.
Clients that retry handle this on their own; some MCP clients don't. If yours doesn't, or you want instant responses,
mark the server always-on — available with a verified phone, billed by the RAM it actually uses.
A typical small MCP server uses well under 256 MB of RAM. Running 128 MB around the clock is about $0.45 a month — inside the $2.25 monthly free credit of a verified account.
Security checklist
- Require a token on every request; keep it in an env file on the server, not in the code.
- Give each tool the least access it needs — a read-only database user, a scoped API key.
- Log tool calls, so you can see what agents actually did.
Try it: paste this to Claude Code, Cursor or Codex.