One binary. Shell in, run commands and copy files by name, through any NAT. No account, no server to host, no config.
A laptop at a cafe. A server behind a home router. No port-forwarding, no public IP.
Your machines talk to each other, end-to-end encrypted, through any NAT. No port-forwarding, no public IP, no VPN appliance.
No account and no coordination server. When a direct path is blocked, a relay forwards encrypted packets it cannot read, and you can run your own with one flag.
Invite a machine and it answers as myserver.hop from anywhere, on a private address only your machines can reach. Your own private network, with ssh, scp and VNC working as they always did.
Close the laptop, switch networks, come back: the shell you left is still there, with its screen intact.
One skill and your AI tools can reach every machine, run a command across all of them, and search their logs together, within limits you set. How agents use it.
Two machines, three steps, under a minute.
Installs WireHop as an always-on service and prints a one-time invite.
Redeem the invite once. The machine is saved under its name.
Just the name. Works from any network, no setup on the router.
macOS (Apple Silicon and Intel) and Linux (x86_64, ARM64, ARMv7). Also on
GitHub Releases as signed binaries and a macOS .pkg,
and on npm as npx -y @wirehop/wirehop.
All install options and how to verify a download.
Every machine you invite is also on a VPN your machines run themselves. We call it a warren.
Each machine gets a private address and a name like myserver.hop, reachable from anywhere by any other machine you invited. No coordination server, no port-forwarding, no config files. The tools you already use just work: screen-share into your desktop over VNC, ssh and scp by name, open a web UI on a box behind a home router.
Bridge a printer, a NAS or a whole LAN through one machine that sits on it. How the private network works.
The same install, invite, connect. More machines, more hands.
Keys never leave your machines. Invites are single-use and time-limited, and every machine is default-deny until you hand one out.
The threat model →Give your agent the skill. It reaches the fleet, runs commands, and stops at whatever you have not allowed.
One skill, no integration →Roles and tags, one command across every machine, scheduled scripts that keep their own state.
From one machine to thousands →