Every scene has a chat panel. Describe what you want and the assistant builds it in the scene you have open, using the same APIs the editor's own panels use — everything it makes is an ordinary object, material, script, or asset you can then edit by hand.

It is not a separate mode. Changes appear in the viewport as they are made, and Cmd/Ctrl + Z reverses them exactly as it reverses your own edits.

What it can build

Scene objects — create, update, and delete, one at a time or in batches. Shapes, characters, cameras, lights, water, GUI elements, soft bodies.
Layout — it plans spatial arrangements and GUI layouts before placing anything, rather than guessing coordinates one object at a time.
Materials and shaders — it reuses a project material that already matches before making a new one, so asking for "a red crate" twice does not leave you with two identical red materials. Shader effects go through a specialist that writes the GLSL.
Lua scripts — object scripts and scene scripts, including attaching a scene script to the scene so it actually receives player actions.
Vehicles — wheeled and tracked, each arriving with a driving script already attached.
Generated assets — noise height maps for terrain, boolean meshes, and swept tracks. It can also read back and edit the recipe of one it made earlier.
Player controls — configuring the scene's input scheme and action buttons.
Ephemeral Lua — it can run throwaway Lua against the open scene to inspect it, the same thing the editor Lua REPL does.

It will ask before guessing

When a request leaves something open that cannot be safely inferred — which of three plausible materials you meant, how big "a large room" is — the assistant stops and puts a short form in the chat rather than picking for you. Questions come with suggested options, and most also accept free text if none of them fit. Answer the form and it carries straight on with the work.

This is deliberate: a wrong guess early in a build is expensive to unpick once forty objects depend on it.

Handing it a broken object

If an object fails to load, the panel that reports it has a Copy button — see When an object fails to load. What it copies is written for the assistant: the object's id, the error, and the object's stored data, so pasting it into the chat gives the assistant everything a repair needs without it having to go looking.

What it cannot do

The assistant works inside one scene at a time — the one you have open. Anything above that level is yours to do in the editor:

Creating, renaming, archiving, or deleting projects.
Creating, deleting, or switching scenes. Ask it to build in a scene by opening that scene first.
Configuring multiplayer settings — player limits, the authority action schema, weapon and entity templates.
Publishing a game or managing releases.
Deleting materials, shaders, or assets. It can create and update them, and it can delete a script, but clearing out unused resources is done in the editor's library panels.
Uploading a file you have on disk. It can generate an asset from parameters or fetch one from a URL, but it cannot reach your filesystem.

How this differs from connecting an MCP client

Both drive the same editor. The difference is where the model runs and how much it already knows.

	The AI assistant	An MCP client
Setup	None. It is in every scene.	Add the server URL and authorize with OAuth.
Scope of work	The one scene you have open.	Every project and scene in the account, including creating and deleting them.
Knowledge of the editor	Built in. It knows the object types, the Lua API, and the conventions without being told.	Whatever the client's model knows, plus the tool descriptions.
Asks clarifying questions	Yes, through a form in the chat.	Depends entirely on the client.
Cost	Counts against your account's usage.	Billed by whoever provides the client's model.

They are not exclusive, and an MCP client is the better tool for work that spans scenes or projects — bulk edits, migrations, generating a dozen variations of a level. See Connect an MCP client.

Getting better results

Say what it is for, not just what it is. "A ramp the player can drive a car up" produces a different ramp from "a ramp".
Build in passes. Ask for the layout, look at it, then ask for the scripting. A single enormous request gives you less to correct mid-flight.
Name things. Once an object has a name you can refer to it later — "make the Crate heavier" works; "make that box heavier" depends on what it thinks you were looking at.
Tell it what went wrong. If something is in the wrong place, say where it should be instead. It can read the scene back, so a correction is cheap.