Setting up the AI.

Lumen works as an ordinary Second Life viewer with none of this done. The assistant stays quiet until you give it somewhere to think.

Open Preferences → AI. The Use dropdown starts on None; pick one of the others. They differ in who pays and in what leaves your machine, and in nothing else — every option reaches the same tools.

Already pay for ChatGPT, Claude or Mistral? Pick that one. Codex (your ChatGPT subscription), Claude Code (your Claude subscription) or Mistral Vibe (your Mistral account) uses the plan you already have: press Set it up for me… and it is done, with no API key to find and no extra bill. An API key is for when you have none of those.

Lumen was created with the help of Anthropic’s Claude, and Claude Sonnet 5.5 is the model that has given the best results. However, you can use any of the options below.

Plans as OpenAI, Anthropic and Mistral listed them in September 2026. They change, so check theirs before paying for anything.

Choosing Codex, Claude Code or Mistral Vibe puts a Set it up for me… button on the panel. It opens a window that does each step for you and checks that it worked before moving on — there is no Terminal to find and nothing to paste.

Codex has three steps, Claude Code and Mistral Vibe two. What it is doing, in plain words: downloading the small helper program that talks to your account, signing it in (in your browser — Lumen never sees that password), and, for Codex, starting it running. After your computer restarts, Lumen starts Codex again by itself.

When it is done the button disappears on its own and a model dropdown takes its place. That is the whole signal that it worked.

Beside the Use dropdown is Test. It sends a real question through the real tools and tells you what came back, so you find out now rather than the first time you ask the assistant something and nothing happens.

For a local model it doubles as a warm-up: the first request loads the model, which can take twenty seconds, and doing that here means your first real question is quick.

Nothing above is required. These are the same commands the window runs, for anyone who would rather see them first. On a Mac, open Terminal; on Windows, open PowerShell. Paste one line at a time.

Codex — install, sign in, start:

Claude Code — install, then sign in:

Mistral Vibe — install by hand, then let the window’s second step sign you in, on Mistral’s own page in your browser:

Then go back to Preferences → AI and press Test.

What you type, and whatever a tool looks up for you — chat, instant messages, the text of a notecard, the names of things you own — is sent to whichever provider you chose, so that it can answer. Their privacy policy is the one that applies. A model running on your own computer sends nothing anywhere. The full list is on the privacy page.

When you talk to the assistant with the mic, only the words it heard go to your provider, exactly as if you had typed them. On a Mac, Apple's speech recognition works out the words: English is done on the Mac itself, while some languages, Danish among them, are sent to Apple to be recognised. On Windows, Whisper does it on your own computer, so the sound never leaves it. Lumen asks before it downloads Whisper's speech model (about 190 MB) the first time you click the mic, and you can remove it again in Preferences → AI → Assistant.

Keys are kept in the same protected store as your Second Life password. That is obfuscation rather than encryption, and the panel says so rather than implying a safety it does not have.

Codex, Claude Code and Mistral Vibe talk to Lumen over a socket on 127.0.0.1 only, never the network, on a port chosen afresh each time the viewer starts. Any request carrying browser headers is refused, which is the attack a local port actually faces. There is nothing to switch on and nothing to configure: it exists only while the viewer is running and only because you picked one of those three.