Your machines, one panel.
This page talks only to an agent on the computer you are using. That agent finds the others. Nothing here leaves your network, and none of it is in the page you downloaded.
Nothing is running here yet
Atlas runs on your hardware, not ours. Install the agent on a machine and this page becomes its control panel.
- 1
Install the launcher
curl -fsSL https://atlasinference.io/install.sh | sh - 2
Start the agent in the background
atlasctl agent install
Already installed? Run them anyway — the first keeps the version you have, and the second starts an agent that is installed but stopped.
Your browser will ask permission to reach other apps on this device. That is this page opening a connection to the agent on 127.0.0.1, and nothing else — it is asked now, rather than on arrival, because until now there was nothing to connect to.
Any web page can show you an install command. Check the address bar says atlasinference.io before running one.
How they reach each other.
Two views: how you reach each machine, and how the machines reach each other. Multi-node decode is all-reduce bound, so the link between two machines decides the throughput. A cluster that falls back to ethernet still runs — several times slower — while every correctness check keeps passing, so the fabric is called out here rather than left to be discovered in a benchmark.
Machines appear here once an agent is running.
Two of them, on the same network, are what the EP=2 recipes need.
Links as text
| From | To | Link | Speed |
|---|---|---|---|
| No links — pair a second machine to form a cluster. | |||
Run one model across them.
Two phases, because one cannot fail cleanly. Every machine checks it can run this and holds its place; nothing starts until all of them have agreed. If one refuses, the reservations the others took are released — so a cluster is either whole or absent, never a half that hangs waiting on a rendezvous.
Connect an agent to launch anything.
What needs looking at.
Idle machines matter as much as busy ones. A clamped clock, a failing fan or a full cache filesystem is something to know before a launch, not after a benchmark comes back wrong.
Nothing to report
No alerts. This section stays here so you never have to wonder where they would appear.