From scene to code, Spline V2 opens its 3D editor to AI agents

Spline V2 adds an agent for creating editable 3D scenes, a local MCP server, HTML/JS code, and a new WebGPU engine.

Spline V2 goes beyond adding a generative feature to the existing editor. The platform has rebuilt its 3D creation environment around a new interface, a WebGPU engine, an integrated coding workspace, and several ways to hand a scene over to an agent.

The built-in agent is located in the editor’s sidebar. A prompt can ask it to create an entire scene or modify an existing project. It begins by reading the file’s structure, uses the same tools as someone working manually, and then captures screenshots of the scene to review the result and correct certain visible issues.

Its scope covers objects, materials, lights, cameras, Boolean operations, particles, cloners, lathed shapes, environments, and variables. It can also build animations, add events, create interactive states, and generate 3D models or images to insert into the scene.

Every action performed by the agent remains a standard editor operation. Changes appear in the undo history, sync with other people working in the file, and can be adjusted manually. The result is therefore not delivered as a flattened render or locked scene.

Users can select objects before submitting a request to restrict the agent’s work to a specific part of the file. Selected references remain attached to the conversation, making it possible to understand later what phrases such as “this object” or “these elements” referred to. Images, media, and documents can also be attached as references.

Multiple prompts can be placed in a queue while the agent is working. The queue can be paused, cleared, or stopped. Conversations are saved on a per-file basis, with the option to maintain several separate threads for the same project.

The documentation acknowledges several limitations. The agent processes only one file and one change at a time. It handles general intent better than exact numerical placement, where manual intervention is often faster. Complex requests may also be completed only partially and require another attempt. No changes can be applied while the scene is running in Preview mode.

Spline V2 does not reserve this automation for its own agent. The Spline MCP server allows Claude Code, Claude Desktop, Cursor, VS Code, ChatGPT, Codex, and other compatible clients to control the editor directly. An external agent can create a file, inspect a scene, move objects, adjust a camera, capture screenshots, or call Spline’s generation features.

The MCP server is built into the desktop app for macOS and Windows. It does not work from the browser-only version. When launched for the first time, the app detects supported clients and adds their configuration automatically. Windsurf and other tools that are not configured automatically require manual setup through Spline’s MCP settings.

Calls are routed locally through `127.0.0.1`. According to Spline, the MCP server does not expose the scene to the network, and exchanges remain between the app and the connected client. Spline’s own generation features still use its online services when called by the agent.

The same connector automatically distinguishes between 3D requests and interface tasks. The former are sent to a Spline tab, while screens, pages, and 2D compositions can be routed to Hana. An agent working inside a code repository can therefore rebuild an interface in Hana, let the user modify it visually, and then apply some of those changes to the original project. This round trip remains limited to the HTML and CSS subset supported by Hana.

The new Code tab also expands the editor’s role. Each file can contain an HTML document combining markup, CSS, and JavaScript. This document runs in a sandboxed environment layered over the 3D scene, with a `spline` API that can read and control its content.

This layer can be used to create an interface, dashboard, game display, or overlaid controls. It can also run without any visible elements to manage animation loops, interaction rules, or game logic. The code can read variables and objects, change their appearance, move elements, create shapes at runtime, and respond to scene events.

The agent can write this HTML/JS document from a description. The resulting code remains visible and editable in the editor, with syntax highlighting, API completion, autosave, and a debugging console. Edit, Code, and Preview modes separate scene editing, document authoring, and execution of the complete experience.

The document stays with the file during collaboration and across several export types, including public URLs, `spline-viewer` integrations, and exports using the Spline runtime. It does not accompany Three.js or react-three-fiber exports, which rely on a different execution method. Some custom Three.js materials are also unsupported in `spline-viewer`.

The rendering engine has been rebuilt around WebGPU, which is now used by default. WebGL remains available as a fallback for unsupported devices or browsers. The transition is intended to reduce the rendering overhead of complex scenes, accelerate their initial display, and support more objects at once.

Some new features depend directly on WebGPU. These include PBR materials, which use roughness, metalness, normal information, occlusion, emission, clearcoat, sheen, or iridescence to describe how a surface behaves under light.

GLTF or GLB files that already use a metallic-roughness structure can be imported with their main textures connected automatically. This continuity is intended to better preserve the appearance of objects created in Blender, Substance, Houdini, or a 3D asset library.

Lighting now includes two types of environments. HDRi mode uses a panoramic image to illuminate the scene and provide reflections. The background can remain hidden while its lighting influence is preserved, a common setup for product presentations.

Physical Sky mode simulates an atmosphere with sun position, time of day, weather, clouds, and wind. Spline also adds screen-space reflections, true planar mirrors, height-based fog, and several material layers. Dust