One agent, three screens.

What a screen is.

A screen is one conversation with an agent. It is a real Claude Code or Codex session, running in a terminal inside Luma World, in your building's project folder.

Every agent has three screens. They share the agent's name, desk, instructions and model, but each one has its own conversation and history. While you talk to the agent on Screen 1, a scheduled task can be running on Screen 2, and the two never get in each other's way.

The screen chips.

Select an agent and look at the top of its Message tab. Under the agent's name is a row of three chips: Screen 1, Screen 2 and Screen 3.

• The highlighted chip is the screen you are looking at. The terminal below belongs to it.
• A chip with a glowing dot and a softly pulsing border is working right now.
• A dim dot means that screen is quiet: finished, waiting, or not started yet.
• Hover over a chip to see its name, for example Screen 2 — working.

Switching and starting a screen

1. Click a chip. The terminal below switches to that screen's conversation.
2. The first time you click Screen 2 or Screen 3, it starts. You choose between starting a new chat and picking up one of the agent's past sessions.
3. Type into the terminal as usual. The screens you are not looking at keep running in the background.

For a hired employee, the row above the terminal also has + New chat, which starts that screen again with an empty conversation, and ⟳ Reload, which restarts the tool and carries on the same conversation. Reload is handy after you change a setting.

The status line

Right under the chips is one line of text about the screen you are on. It is always there, even when it is empty, so the panel does not jump around.

• While the agent works, it shows what the agent says it is doing, with a timer and a count of tokens used.
• When the agent finishes a task, the line turns green and starts with a ✓, followed by the agent's one-sentence recap.
• Otherwise it shows the last thing that screen reported.

+ New chat clears the line. ⟳ Reload keeps it.

Monitors on the desk.

Each screen you have started gets a monitor on the agent's desk. An agent using only Screen 1 has one monitor. Start Screen 2 and a second monitor swings in beside it.

Each monitor shows what its screen is doing right now: the file being read or edited, the command being run, or a question waiting for your answer. Between tasks it shows a quiet "connected" screen with the agent's name. The monitor for the screen you are viewing in the panel is the bright one; the others are dimmed but keep updating.

Claude Code or Codex.

An agent runs on one of two tools: Claude Code or Codex. You choose by choosing the agent's model. Claude models run on Claude Code; GPT models run on Codex.

All three of an agent's screens use the agent's model, so they all run on the same tool. There is no separate runtime switch per screen. To change the model:

1. Select the agent and open its Settings tab.
2. In the Model section, pick a model. The one in use is marked current.
3. On Claude Code, Screen 1 switches straight away: the app types a /model command into its terminal, so the running conversation carries on with the new model. The other screens keep their model until their next session starts, for example after + New chat or ⟳ Reload.
4. If you picked a model from the other tool, every screen needs a new session to move over, Screen 1 included. Each new session starts on the new tool.

You can also set the model when you add an agent: the New Agent dialog has one row of model buttons for Claude Code and one for Codex.

Your own subscription comes first.

Agents run on the Claude Code or Codex you already have, signed in with the plan you already pay for. Luma World does not charge for them, and with a sign-in they need no API key.

• Claude Code uses your Claude sign-in on this computer, exactly as if you typed claude in a terminal yourself.
• Codex uses your ChatGPT sign-in from codex login.
• No subscription? If the tool is installed but not signed in, an API key can stand in: an Anthropic key for Claude Code, an OpenAI key for Codex. Your sign-in always comes first; the key is only used when there is none. See Using an API key instead.

Some of Luma World's own features also use AI, for example speaking to an agent, the Architect's voice, or the seats in a Circle meeting. For those, the app again tries your subscriptions first. An API key is only needed to fill a gap that no subscription covers.

Where to check

Open Settings and go to the AI tab. At the top, Agent runtimes shows whether Claude Code and Codex are Ready, Needs sign-in or key or Not installed, with a Check again button. A tool that is not ready has an Install Claude Code or Install Codex button (or Sign in once it is installed), and a command you can paste into a terminal instead. Below that, each kind of AI work (Realtime, Language model, Voice to text, Text to speech, Embeddings) shows whether a credential was found. Under Your keys you can paste an OpenAI, Anthropic or other key. Keys stay on your computer.

When a screen will not start.

If a screen cannot start, you do not get a blank terminal. You get a card that says what went wrong and what to press next.

The tool is not installed

You see Claude Code isn't installed (or Codex isn't installed), with a command you could paste into a terminal and a Copy button. The easiest fix is inside the app:

1. Press Open Settings → AI, and under Agent runtimes press Install Claude Code or Install Codex. The app installs it for you, the same way as the setup screen.
2. Press Sign in there once it is installed, and finish in your browser. If you have no subscription, add an API key under Your keys instead.
3. Come back and press Check again. If the tool is there now, the chat starts.

If it still says Still not found, open a new terminal window once, then check again. The card also offers Back to sessions. A hired employee's list of past sessions shows the same install help while its tool is missing.

The chat could not start

For other problems the card reads Could not start the chat (or Could not start Codex), with the reason underneath and these buttons:

• Try again starts the screen again.
• Back to sessions returns to the list of past sessions, so you can pick a different one or start fresh.
• Open Settings appears when the fix is in Settings → AI.

If a chat's tool stops, the terminal ends with a grey line such as [process exited with code 1], or [stopped by another process (SIGKILL)] when something outside the app stopped it. When the exit was an error, the session list says The last chat closed with an error (exit code 1). Start a new chat or pick a past one., with a Dismiss button.

The hidden Circle screen.

When you invite an agent to a Circle meeting, it prepares on a fourth, hidden screen, so the meeting never types into the conversation you are having with it.

You will not see this screen among the chips or on the desk. Instead, a Circle tab appears in the agent's panel, right after Dream, while the agent is preparing. Open it to watch the agent read your project and write its notes for the meeting. The tab's status line reads, for example, Preparing · 4:19 left, and turns amber while the meeting is waiting on this agent.

When the meeting starts, the agent's seat speaks from the notes it wrote, and the Circle screen closes by itself a few seconds later. Screens 1 to 3 carry on the whole time.

Common questions.