Spline V2: AI-agent collects interactive 3D scenes, and the interface can be written on top of them

On August 20, 2026, Spline released V2 – an updated 3D editor with an AI agent, rendering on WebGPU and HTML/JS layer on top of the scene. For a designer, this is an opportunity to collect a three-dimensional object, its behavior and control interface in one file: for example, a promo block with a choice of color or a small game.

We analyze the August release according to the official announcement and documentation, checked on September 8. The main interest here is the transition from a beautiful 3D object to a scene that the visitor can interact with.

What does an agent do

Built-in AI Agent works through editor commands: creates objects, adjusts materials and light, adds animation and connects interactivity. Its changes fall into the cancellation history and are synchronized with the participants in the file. The result can be further edited manually.

The Agent panel is on the left side of the editor. Before the request, you can single out the object: then the edit “make it matte” gets a specific goal. In the documentation, Spline recommends consistent refinements instead of one huge task.

The agent has limitations: it works with one file and one edit at a time, and the exact coordinates are sometimes faster to set yourself. A difficult task can be partially completed. Therefore, after a change in behavior, it is important to see what really happened, rather than limiting yourself to a chat response.

Source: aI Agent Documentation.

Why you need HTML/JS over 3D

V2 has three modes: Edit, Code and Preview. Code stores an HTML document with CSS and JavaScript, and Preview shows it along with the scene. Through the built-in API, code can access Spline objects and variables.

So you can make a material selection panel, a point counter or an animation control button. The interface and the volume part remain connected within the same file.

There is an important nuance of export: the HTML layer is transferred to the Public URL, embedding through a spline-viewer and exporting with the Spline engine, but not to export three.js/react-three-fiber. In addition, changes to objects from executable code are a temporary state: they are reset when rebooted and do not turn into saved edits to the original scene.

This affects the choice of the method of embedding. If the prototype relies on Code, the export format should be checked before the entire site is built around it.

Source: Code Tab and export restrictions.

What's changed on the schedule

Spline has moved the renderer to WebGPU, keeping WebGL as a backup option. The update also includes PBR materials, HDR lighting, reflections and new modeling tools.

The company reports that it is accelerating its work with complex scenes and reducing the size of exports. These statements describe improvements to the engine, but do not give a ready-made forecast for each landing: the speed of a particular promo block still depends on its content and the visitor’s device.

Source: official announcement of Spline V2.

You can work through an external agent

Spline MCP connects compatible AI tools to an open source editor. Such a connection requires a Spline desktop application for Windows or macOS: a browser tab is not enough.

After installing and running the application, the supported AI client may need to restart. Connections are managed in Settings → MCP. The external agent receives tools to read and change the scene, create objects and work with the code layer.

Source: Spline MCP Server.

Example: Promoblock with color selection

We offer a small scenario for self-checking. This is an editorial example of a task, not a test report.

First, ask the agent to assemble the base:

Create a 3D scene for a table lamp promo block. The lamp stands on a low podium, the background is light, the lighting is soft studio. Body matte, no logos. Leave a free seat on the left under the title. Name the objects so that the body, podium and light source are easy to find.

After the visual check, add the behavior:

Make an HTML panel with three buttons: Graphite, Coral and Sand. At the press, change only the color of the lamp body. The selected button should be visible. Add the spin switch and reset button. The catwalk, camera and lighting should be maintained.

Separately set the conditions for the mobile screen:

Check the composition on a narrow screen. The panel should not block the lamp or block interaction with the entire scene. Buttons should work from the keyboard. When the reduced motion preference is turned on, the automatic rotation should be turned off.

When checking, go through the entire script: change color several times, stop rotation, reset the state and reboot the preview. Then check the same scenario in the selected export format. The working result is not only a successful angle, but also predictable reactions to the actions of the visitor.

What tasks are worth trying for

The most understandable uses are interactive product demonstrations explaining scenes and small game elements in promotional projects. They can limit the number of objects and actions, quickly test the idea and decide whether the 3D complexity of the page justifies.

The built-in agent spends AI workspace credits; tariff conditions should be checked before a long series of generation. For the first acquaintance, one scene with two or three actions is enough: this volume allows you to evaluate both the visual result and the controllability of subsequent edits.