Editing Worlds with an AI Agent

Ask your browser's AI agent to "line up ten chairs here" and the objects appear in your XRift world. This works for worlds you can edit in the browser (EDITOR type), and is built on WebMCP.

Whatever the agent places shows up immediately for everyone in the instance and is saved to the world. It is treated exactly like something you placed by hand, so you can undo it with Cmd/Ctrl+Z and keep editing it from the inspector.

What is WebMCP?

A page registers a list of "things you can do here", and an AI agent built into the browser can call them. Instead of clicking around the screen, the agent calls the functions the page exposes directly, so it does not depend on how the UI looks.

XRift exposes its world editing operations this way. If your agent supports WebMCP, the tools are visible without installing anything.

Requirements

Browser	Chrome 149 or later
Site	https://app.xrift.net
World	A world editable in the browser (EDITOR type)
Permission	Owner of the world, or a user granted edit access

Enter an instance, and if the edit button (🔧) is visible you are ready. You do not need to enter edit mode — you can ask for a wall while walking around.

note

This ships through Chrome's Origin Trial, so it works on app.xrift.net as-is until November 17, 2026. After that, further work may be needed depending on how the specification is finalized.

What it can do

These are the operations the agent can see. You never need to name them yourself — describe what you want and the agent picks the right ones.

Tool	What it does
list-placeable-types	Look up placeable types and their dimension ranges
get-scene	Look up what is currently placed in the world
get-viewer-context	Look up where you are looking and how many objects can still be added
place-objects	Place objects in bulk (up to 50 per call)
update-objects	Change position, rotation, scale, dimensions, color, material
remove-objects	Remove objects (groups are removed with their contents)
group-objects	Combine several objects into one group
undo-last-agent-edit	Undo what the AI just did

Five types can be placed: box, floor, wall, sphere, and cylinder. Image panels and screen share displays are excluded, because we do not want an AI inventing URLs or stream sources. Spawn points are also left to humans, since they change where everyone enters.

Tips for asking

Say what the reference point is. "At the world origin" uses absolute coordinates; "where I'm looking right now" uses the surface under your crosshair. The latter is usually easier to describe.

Build a 2m square stage on the floor in front of me

Use meters. One unit is one meter in XRift. Real dimensions like "a 3m wall" or "a sphere with a 50cm radius" pass straight through.

Ask for things in bulk. "Five chairs, 2m apart" is placed in a single call. That is faster than one at a time, and it undoes in one step.

Ask for a group when you are done. "Group the bench you just made" turns it into a single object you can move or delete as a unit.

Roll it back freely. "Undo that" reverts it, same as Cmd/Ctrl+Z.

Seeing what happened

When the agent places something, a banner appears at the top of the screen.

• 選択 (Select) — selects what was placed (and enters edit mode), for fine-tuning
• 元に戻す (Undo) — reverts it, same as Cmd/Ctrl+Z
• ✕ — dismisses the banner

The banner hides itself after 10 seconds, but you can still undo after it is gone.

Your input always wins

The AI never takes over what you are doing.

• It does not steal your selection — if you are editing something in the inspector, your selection stays put even when the AI places objects
• It does not switch modes — edit mode stays however you left it
• It backs off while you are working — requests are refused while you are placing an object, in the middle of a move (M), or answering a delete confirmation dialog
• It leaves other people's work alone — objects someone else has selected are excluded from changes and deletions, and reported back as "N object(s) being edited by someone else were left alone"

Troubleshooting

The agent cannot see the tools

1. Are you on https://app.xrift.net in Chrome 149 or later?
2. Are you inside an instance? (Nothing is registered on list screens.)
3. Is the edit button (🔧) visible? If not, you lack edit permission for that world, or it is a code-built world (CODE type) that cannot be edited in the browser.

"The object limit has been reached"

A world can store up to 500 objects. get-viewer-context reports the remaining capacity. Remove what you do not need and ask again.

Objects land on the wrong side

The reference for "in front of me" may not have come across. Naming what you are looking at — "on the wall I'm facing" — makes it reliable.

You cannot find what was placed

Press 選択 (Select) on the banner: it enters edit mode with the new objects selected. This also finds objects that ended up below the ground or far away.

For developers

Tool input/output schemas, the coordinate and rotation conventions, and the pitfalls to avoid when implementing an agent are collected in the xrift-world-editing skill in xrift-skills, formatted for AI coding agents.

npx skills add WebXR-JP/xrift-skills