Architecting AI-Powered Collaborative 3D Editing: Spline's MCP Server
Spline's V2 rebuild introduces the MCP Server, a local component that allows external coding agents like Claude Code to directly manipulate live, editable 3D scenes. This architecture enables a seamless feedback loop between AI agents and visual designers by treating AI-driven changes as standard editor operations, integrating them into the undo history and collaborative sync mechanisms. The article highlights the design of a local server for secure, low-latency interaction and the automatic handoff of tasks between 3D editing and UI components.
Read original on The New StackIntroduction to the MCP Server Architecture
Spline V2 introduces a significant architectural shift by integrating a local server, the Spline MCP Server, directly into its desktop application. This server acts as a bridge, allowing external AI coding agents (e.g., Claude Code, Cursor) to interact with and modify live 3D scenes. This design pattern is crucial for enabling real-time, iterative changes to complex visual projects, moving beyond static image generation to dynamic, editable scene manipulation.
Key Architectural Components and Interactions
- Spline MCP Server: A local server bundled with the desktop application (macOS, Windows). It's bound to 127.0.0.1 for security and uses an origin allowlist.
- External AI Clients: Tools like Claude Code, Cursor, Codex, and VS Code that connect to the MCP Server. They send structured MCP tool calls based on user prompts.
- Live 3D Editor: The core Spline application where changes are rendered and applied in real-time. It processes MCP calls as standard editor operations, ensuring editability and integration with undo/redo history.
- Hana (UI Editor): A separate component for interface changes, also accessible via MCP, allowing agents to modify both 3D scenes and UI elements.
The design emphasizes local processing for responsiveness and security. AI agent requests are translated into structured MCP calls that hit live scenes. This allows the agent to 'see' the current state of the scene before making changes, facilitating a coherent back-and-forth between visual design and code.
Decoupling AI Logic from Core Editor
The MCP Server approach effectively decouples the AI agent's logic from the core editor. The editor provides a stable API (MCP) for modifications, treating AI input like any human-driven edit. This ensures maintainability, allows for diverse AI tool integration, and preserves the editor's core functionality (e.g., undo history, collaborative sync).
Security and Data Flow Considerations
Security is a primary concern. The MCP server operates locally, bound to `127.0.0.1`, and uses an origin allowlist to restrict client access. Prompts and scene data exchanged with the MCP server remain on the local machine. However, if the AI agent utilizes Spline's own AI generation features (e.g., for new models or images), those specific tasks might still involve cloud processing, similar to manual invocation from the editor. This distinction highlights a hybrid architecture for AI-driven features.
For collaboration, the article notes that multiple agent sessions can connect to the same editor, but it doesn't detail how attribution or review of individual MCP changes are managed before publishing shared files. This points to a potential area for further architectural development in multi-agent, collaborative environments.