Urgent.News

What's breaking now, across thousands of outlets.

Editions

AI

Spline rebuilt its entire 3D editor. Then it handed the keys to Claude Code.

Spline released V2 on Thursday, a complete rebuild of its 3D editor that enables external coding agents to work directly The post Spline rebuilt its entire 3D editor. Then it handed the keys to Claude Code. appeared first on The New Stack .

Spline rebuilt its entire 3D editor. Then it handed the keys to Claude Code.

Spline introduced V2, a complete overhaul of its 3D editor, on Thursday. This new version allows external coding agents to interact directly with live, editable scenes. A crucial addition, the Spline MCP Server, enables tools like Claude Code, Cursor, Codex, Google Antigravity, and VS Code to function within a live Spline project.

Through a bundled local server with the desktop app, these tools can modify the scene and its behavior while preserving the project's full editability. The MCP server, integrated directly into Spline's rebuilt desktop application for macOS and Windows, does not function in the browser and requires no standalone installation. When the desktop app launches, it identifies compatible clients and registers the server in their configuration files.

After the developer restarts the AI client, they can issue commands normally. For instance, if Claude Code is asked to create a floating island, it translates the request into structured MCP tool calls. Spline then directs these calls to an open 3D editor tab, where they implement changes against the live document. If no suitable file is open, the application creates one.

The model can view the scene's existing state before making alterations, enabling it to navigate between visual design and code running on the canvas, utilize Spline's AI tools as needed, and continue editing the interface or game logic. All changes remain within the Spline project, ensuring the designer receives the live scene back post-agent completion.

Spline's internal AI agent operates similarly, treating each change as a standard editor operation visible in the undo history and syncing across collaborators. Spline isn't alone in granting MCP keys to external agents; ElevenLabs recently opened its voice-agent infrastructure to Claude via an MCP server, allowing a chat window to manage production voice agents similarly to how Spline enables coding agents to manage 3D scenes.

Spline's approach allows agents to traverse the canvas gap, as V2 unifies the 3D editor and Hana within the same MCP connection, automatically routing scene work to the 3D editor and interface changes to Hana. In Hana, agents can transform components into editable frames and carry visual changes back into application code, using stable references on exported elements to locate the correct place in the source code.

This round trip has limitations due to Hana's support for only a subset of HTML and CSS based on flexbox, potentially requiring reconciliation of complex markup and framework-specific components with existing code. For 3D work, developers can build or modify a scene using Claude Code, refine it visually in Spline, and export it for the web in formats such as Vanilla JavaScript, Three.js, React, Next.js, and React Three Fiber.

These exports default to WebGPU, falling back to WebGL when necessary. Spline's runtime API can control a published scene, but the company does not document a method for changes made to the exported code to flow back into the original 3D file. If an agent needs to revise that source, it must return through MCP and operate in the live editor.

Google Antigravity, recently expanded beyond its own IDE with extensions for VS Code and JetBrains, now has another surface to work on, thanks to Spline. OpenAI's Codex, another supported client, has recently gained the ability to continue coding while awaiting a developer's response. Editing before export is significant, as Spline already provides a Code API that allows a web application to control the state and behavior of a published scene in real-time while it is being built, granting the agent direct access to the editable document to comprehend its structure and make changes before export.

Here, MCP acts as a control interface for the editor, while Spline manages the visual project and renders changes as they occur. The connection remains local, bound to 127.0.0.1 and protected by an origin allowlist that prevents arbitrary web pages from accessing the server. Developers can specify which clients have access through Spline's MCP settings.

Spline reveals that prompts and scene data exchanged with its MCP server remain between the desktop app and the connected client. However, if the agent employs one of Spline's AI generation features, this work continues in the cloud, just as it would if initiated manually from the editor. Multiple agent sessions can connect to the same editor, but Spline's documentation does not clarify how teams can attribute or review individual MCP changes before publishing a shared file.

Written by urgent.news from The New Stack's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at thenewstack.io →

More in AI

Nvidia just showed that the harness, not the AI model, is now the real hero

Nvidia research shows that AI agents can perform well, and not go off the deep end, through fine-tuning, even if the AI model isn't that great at the task.

  • Nvidia's research reveals harness, not AI model, is crucial for long-horizon tasks.
  • Optimized harness and supervisor component boost Claude Opus 5's ARC-AGI-3 score to 100%.
  • Harness includes memory, context, feedback, and scaffolding for AI agents.

AI Boosted Homework Scores, Then Exam Scores Dropped

An anonymous reader quotes a report from The Economist: Students and school children are increasingly using artificial intelligence.

  • AI usage by 80% of Chinese students aged 12-18
  • Homework scores increased by 18% with AI assistance
  • Exam scores dropped by 20% with AI usage

More from Friday 21 August →