A room that shows you the work.

Globe, city, building.

There are three places between opening the app and watching your agents. Each is a 3D view with a panel on top, and everything you can click in 3D you can also pick from the panel.

1. The globe. A planet with your cities on it. The Choose a City card lists them. Click one, or press ＋ New City, type a name and press Create.
2. The city. A row of temples. Each temple is a building. When you add one, the Building Details dialog asks for a Title, a Handle, an optional Description, and a Working Directory with a Browse button.
3. The building. Click a temple to walk in. If the building has no folder yet, the app asks you for one first.

To go back, use the toolbar: City leaves a building, and Cities returns to the globe. Press ⌘ R from anywhere to jump straight to another building.

Walking in

While the office loads you see entering and the building's name. Then a short welcome plays: the building's name settles letter by letter under — Location —, and a voice says "Welcome to" your building. It ends on its own after a few seconds, or you can Click to skip.

The 3D office.

The office is a small 3D room with desks. The room only draws the agents that are really in it, and it updates as they work. You can drag to turn the view and scroll to zoom.

Floors and rooms

A building can have more than one floor. The floors are listed in the rail on the left, under Floors.

• Switch floors by clicking a floor in the list.
• Add a floor with + Floor.
• Rename a floor by double-clicking it. You can also drag floors, or use Alt + ↑/↓, to reorder them.

Under This floor, the Room switch picks which room this floor shows: ⛩ Zen (The Zen Garden), ⬡ Holo (The Hologram Office) or ◆ Studio (The Studio). On a Zen Garden floor, the Monitors switch changes the desk monitors between ◇ Glass and ▭ Bezel. The app remembers which floor you last looked at in each building.

Moving an agent to another floor

1. Select the agent and open its Settings tab.
2. Open the Floor section and pick a floor from the list.
3. For a hired employee you see Saved for future sessions. For a temporary agent, For this session.

Moving an agent moves only its character. Its conversations carry on exactly as they were.

Desks and agents

With nobody selected, the card in the corner reads Crew with a count of your agents, and has a + Agent button. Press it to add an agent:

1. The New Agent dialog shows a portrait and a suggested name. Press ⟳ for another name and face, or type your own.
2. Optionally write an Initial message: what should they start on?
3. Pick a Model. Claude models run on Claude Code, GPT models on Codex. If one of the two is not set up on this computer, its models are greyed out under a note such as ⚠ Codex isn't installed, with a button that fixes it. See When one tool is missing.
4. Press Confirm. The character appears at a free desk, and you hear "New agent has joined."

A new agent is temporary: it lasts as long as its session. To keep it, select it and press ★ Hire. A hired employee keeps its name, desk, floor, voice and memory between sessions. A temporary agent has a Remove button; an employee has Fire Employee.

What a character tells you.

Each character is a small origami figure called a Luma. At a glance it tells you three things: whether its agent is working, what the agent says it is doing, and when it has finished.

A Luma's hat, robe and colours come from the agent's name, so the same name always gets the same look, on any computer. The small portrait in the panel matches it.

The line over its head

Over each character floats its name and a short status line, such as "Writing tests". The agent writes that line itself, in its own words, to tell you what it is working on. Ordinary activity like "Reading config.ts" is not put here; you see that on the desk monitors and in the panel instead.

If the conversation has a title, the line reads as the title, then an arrow, then the status. The whole thing is kept short (about 80 characters). The line clears when the turn ends or the agent finishes.

When it has a question

When an agent stops to ask you something, you will notice in several places at once:

• The character turns to face you and waves.
• The question appears over its head.
• Its desk monitor reads Waiting for your answer, with the options listed.
• You hear "This agent has a question." If nobody answers for a while, "This agent is still waiting for an answer."
• In the agent list, its row reads Waiting for your answer and moves to the top.
• For an agent on Claude Code, the question also appears in the panel as a card under the terminal, just above the Speak bar, with the options numbered. Click an option to answer. If the question lets you pick several options, or asks more than one thing, the card sends you to the terminal to answer there instead.

When it is done

When an agent finishes a task, it files a one-sentence recap. Then:

• You hear "Task complete.", then "Name completed their task." followed by the recap.
• A Task Complete card drops in over the office, with the agent's face, name and recap. Up to four stack at once, and each one leaves after 15 seconds, or when you press ×.
• The agent's status line in the panel turns green with a ✓ and the recap.
• If the panel is hidden, the compact dock on the right shows ✓ and the recap under the agent's name.
• A soft chime plays once the agent has nothing else running.

Selecting an agent.

Click a character in the room, or click its name in the agent panel's list. Both select the same agent.

• A glowing ring appears on the floor under the selected character.
• The agent panel switches to that agent's conversation.
• If the panel is hidden and you click a character, its chat opens in a small window floating above its head.
• Press Esc, or click an empty patch of floor, to deselect.

Selecting an agent does not move the camera. The camera only moves by itself during a voice conversation, when it turns to whoever you are talking to. Drag, scroll or press Esc to take it back.

The agent panel.

The window has two halves. On the left is the office and a rail of tools. On the right is the agent panel: a list of your agents, and the conversation of whichever one is selected.

The rail on the far left is a column of icons; hover over one to see its name. From the top: Office, then Architect, The Circle, Board, Files, Browser, Canvas, Git, Memory Palace and Messages. The Architect and the Circle open full width and hide the agent panel while you use them.

The tabs

Across the top of the selected agent's conversation is a row of tabs:

Talking to one agent

The conversation on the Message tab is a real terminal. Click into it, type, and press Enter. To speak instead, press Speak under the terminal, talk, and press Stop. Your words are typed into the terminal but not sent, so you can read and fix them first, then press Enter.

The agent list

At the top of the list is a search box (⌘ K) and four filters, each with a count: All, Needs you, Working and Idle. Needs you pulses while any agent is waiting for an answer. Agents running now are listed under Live, with the most urgent first: waiting for you, then waiting on helpers, then working. Hired employees who are not running are under Recent and Offline. Press + New to start a new agent from the list.

Around the edges

• The compact dock. When the panel is hidden, a slim strip on the right lists who is online and offline. Click an online agent to open its conversation.
• The status bar along the bottom shows the project name, the git branch, whether you have uncommitted files, and the app's version. On the right, a face for each agent that is busy right now, with its current line: up to four, then a +N button for the rest. Click a face to select that agent.
• Task Complete cards appear over the office when an agent finishes.

Talking to the office.

At the bottom centre of the office is the task box. Use it to hand a job to the right person without picking them first.

By typing

1. Type the job into the box: Describe a job and pick who takes it…
2. The people who best fit the job stand up, with what they are good at on show. Anyone whose tool is not set up on this computer is greyed out and marked, for example, Codex isn't installed.
3. Click one. The job goes to a free screen of theirs.

By voice

1. Press Talk in the task box.
2. Say "hey" and the name of an employee, then describe the job.
3. The camera turns to the person you are talking to. When you confirm, the job goes to a fresh screen of theirs.

The box shows short receipts as it goes: Handed over, Starting up, Working on it, Finished, Couldn't start or Not sure yet. If nobody matches the name you said, a receptionist voice tells you, for example that the person is on a different floor, or asks which of two similar names you meant.

If the call is quiet for a while, the box warns you that voice will stop in 30 seconds, with a Keep call open button. Your draft is kept either way.

Common questions.