MCP AI Alpha

Connect your local AI agent to WorldLoop CAD.

The standard WorldLoop workflow still uses the Roblox Creator Store connector for Studio delivery. MCP AI is a separate alpha layer for advanced testers who want a local AI agent to inspect, generate or edit WorldLoop CAD context.

Alpha status: interfaces, setup steps and supported edit operations can change while the agent workflow is hardened.

What MCP adds

MCP lets a trusted local AI client read the current WorldLoop workspace context, use the CAD generation catalog, inspect the floor-plan report and submit an edited build back to the web UI.

It is not required for normal generation. Generate, preview and send-to-Studio continue to work through the web dashboard and the Creator Store connector.

Current architecture

  • Standard Studio delivery: the web app queues a build in the cloud, and the WorldLoop Creator Store connector receives the manifest inside Roblox Studio.
  • MCP AI alpha: the optional MCP Helper exposes a local mailbox for an MCP-capable AI client. The agent reads the request, produces a reviewed definition and submits the result back to the web page.
  • Roblox Studio output: after the edited build is ready in the web UI, you still send the final build to Studio through the Creator Store connector.

How to connect an agent

  1. Open the WorldLoop builder and switch to the MCP AI workspace.
  2. Generate or sync the current context so the page sends `aiCadGenerationCatalog`, `cadPlanReport` and `aiPlanContext` to the relay.
  3. Install or check MCP Helper when the page asks for it. This is separate from the Creator Store connector.
  4. Configure your MCP-capable local AI client to run the WorldLoop Studio adapter.
  5. Ask the local agent to call `worldloop_get_page_request`, perform the edit or generation, then call `worldloop_submit_page_result`.
  6. Review the preview on the page, then use `Send Edited Build` to deliver the final output to Roblox Studio.

Advanced testers using the repository can start the adapter with `npm run mcp:studio` or test one queued request with `npm run mcp:page-request:once -- --prompt "make the interior plan playable"`.

MCP client example

Point your local MCP client at the adapter script. The exact path depends on where your WorldLoop checkout or MCP package lives.

{
  "mcpServers": {
    "worldloop-studio": {
      "command": "node",
      "args": [
        "C:\\path\\to\\worldloop\\tools\\mcp\\worldloop-studio-adapter.js"
      ]
    }
  }
}

Rules every agent must follow

  • Use the WorldLoop CAD catalog and fixture vocabulary as the source of truth.
  • Keep real playable scale for floors, rooms, doors, windows, stairs and landings.
  • After every generation or edit, rebuild or read `cadPlanReport` and `aiPlanContext`.
  • Inspect `issues`, `repairHints`, `blockedZones`, stairs, openings and rooms before submitting.
  • Repair supported errors or warnings. If a conflict cannot be fixed, report it instead of marking the preview valid.
  • Use smart mesh as a visual layer, not as the structural source of truth.

Alpha limits

MCP AI currently supports controlled generation and edit operations, but it is not a fully autonomous architect yet. Some prompts may produce partial edits, conservative fallbacks or validation warnings.

Only connect agents you trust. Keep the MCP Helper local to your machine, do not expose its local port publicly, and review generated output before publishing any Roblox experience.