Getting Started

OpenTunnel gives your AI agent the ability to run commands on a remote machine as if they were local tool calls. This page takes you from nothing to a working tunnel in about a minute. No installation, no account, no configuration.

Your first tunnel

1. Start a session on the machine your agent should reach.

   On the remote machine run:

   curl -fsSL https://opentunnel.sh | sh

   This downloads a temporary CLI from the relay, verifies its checksum, and opens one foreground session. The process prints a ready-made prompt for your agent and keeps running. As long as it runs, the tunnel is open; the moment it stops, access is gone.

   The domain answers command-line clients directly with a short, self-documenting script. Prefer a more explicit URL? curl -fsSL https://opentunnel.sh/create | sh serves the identical script. Curl either URL without the pipe first to read what runs.

2. Paste the printed prompt into your agent.

   The session prints something like:

   I opened an OpenTunnel session for you.
   
   Run commands on my host with:
   
   curl -fsSL https://opentunnel.sh/cli | OPENTUNNEL_INVITE='<invite>' sh -s -- exec \
     -- '<COMMAND>'

   Hand that to your agent: OpenCode, or anything that can run shell commands. The agent now has a remote shell as a tool: it substitutes its command, runs the line, and receives stdout, stderr, and the real exit code, exactly as if the command had run locally.

3. Let the agent work.

   From here you talk to your agent as usual. It handles the tunnel on its own and runs whatever the task needs on the remote machine. For example:

   Install a Minecraft Crossplay server.

   Commands execute one at a time, without per-command approval, for as long as your foreground session is running.

4. Press Ctrl+C when you’re done.

   The session ends, the invite expires, and the relay forgets the connection ever existed. There is nothing to clean up and nothing left behind.

Invites are bearer secrets

Anyone holding a valid invite can attempt to connect while your session is active. Don’t paste invites into shared logs, tickets, or long-lived notes. Treat them like a password that happens to expire when you hit Ctrl+C.

Good to know

• Commands must be non-interactive: there is no PTY and no stdin in v1. Avoid sudo unless it is passwordless.
• One agent, one command at a time. Only a single agent can connect to a tunnel; multiple agents per tunnel are not supported, by design. This is part of the security model, not a limitation to engineer around.
• The temporary CLI is cached in the system temp directory for the duration of the session and verified by checksum on every cache hit.

Prefer your own relay?

Everything above works identically against a relay you run yourself: one stateless container, no database, no accounts. Your agents then fetch from your origin instead of opentunnel.sh:

curl -fsSL https://relay.example.com/cli | sh -s -- create

See the self-hosting guide; it genuinely is one command.