Spline V2 represents a significant architectural pivot in how creative tools interact with generative AI agents. The release introduces a bundled local MCP (Model Context Protocol) server that exposes live, editable scenes to external coding clients like Claude Code and Cursor. Unlike previous iterations where an agent might generate static assets or require manual handoffs between chat windows and design files, this implementation allows the model context protocol calls to hit open 3D editor tabs directly within a desktop application on macOS or Windows.
Architecture of Live Scene Manipulation
The core change is the decoupling of scene state from static exports. When an agent issues commands, such as creating a floating island, it translates these requests into structured MCP tool calls routed to live documents rather than generating flattened images or meshes.
This architecture relies on two distinct components: Spline's internal AI agent and the external coding agents connected via MCP. The system handles handoffs automatically between visual design work in the 3D editor and interface changes managed by Hana, a component builder behind the same connection point. However, practitioners must note that this round-trip capability has limits; while stable references allow carrying visual changes back into application code for supported HTML/CSS subsets (specifically flexbox), complex markup or framework-specific components may require reconciliation rather than direct replacement.
Operational Implications and Export Flows
The operational model shifts from a linear build-to-export process to an iterative loop. Agents can now refine scenes visually in Spline before exporting for web formats like Vanilla JavaScript, Three.js, or React Three Fiber using WebGPU by default.
Crucially, the runtime API allows control of published scenes while running, but changes made via MCP during development do not automatically flow back into exported code. If an agent needs to revise source logic after export, it must return through the MCP connection and work within the live editor again. This distinction is vital for platform engineers managing CI/CD pipelines that rely on static artifacts; dynamic scene updates require a persistent connection to the local server.
Security Considerations
The implementation prioritizes scope isolation by binding the local MCP server strictly to 127.0.0.1 and enforcing an origin allowlist to block arbitrary web pages access. Prompts and scene data exchanged between the desktop app and connected clients remain on-premise, though AI generation features still execute in the cloud.
Security teams should evaluate how multiple agent sessions connect simultaneously to a single editor instance without documented mechanisms for attributing or reviewing individual MCP changes before publishing shared files. This lack of audit trails represents an operational risk when managing collaborative environments where distinct agents modify live state concurrently.
Related CloudNinjas coverage: AI engineering.
