Skip to content

Private prototype · public capability face

LLM as controller. Editor as device.

ReVerse is a custom MCP server that lets Claude Code drive UEFN like a CLI — natural-language scene direction becomes typed tool calls with real side effects in the editor. The model decides; the protocol enforces shape.

control loop dry-run ready
01

Agent

Claude Code plans in natural language and selects tools. No raw editor API from the model.

02

ReVerse MCP

Python server: schemas, validation, session hygiene. Side effects only through declared tools.

03

Editor bridge

UEFN listener applies approved ops on the editor thread. Actors move. Internals stay private.

MCPTool protocol
UEFNEditor target
TypedSide-effect surface
NL → opsControl loop

Architecture · public shape only

Three layers. No magic in the middle.

What you can know without seeing the private repo: intent stays in the agent; effects stay in the editor; the bridge is a narrow, typed contract.

01 · Agent

Claude Code (or any MCP client) plans in natural language and selects tools. No direct editor API from the model.

02 · ReVerse MCP

Python MCP server: tool schemas, validation, session hygiene. Side effects only through declared tools — not freeform shell.

03 · Editor bridge

A listener inside UEFN applies approved ops on the editor thread. You see actors move. You don’t get a dump of internals here.

Interactive demo · simulated session

Direction in. Receipts out.

This is a dry-run of the control loop — not a live link to your editor. Tool names below are capability-shaped (place / query / validate), not a leak of the private catalog.

agent · intent
tool stream · viewport IDLE
scene emptyops 0

Capability surface · abstract

What the tools are for — not how they work inside.

Enough for a peer to understand the product. Not enough to clone the private measurement layer or asset graph.

PL

Place

Spawn actors from an approved catalog reference. Deterministic transforms. No “vibes-based” coordinates from the model alone.

XF

Transform

Move, rotate, and scale with editor-safe conventions. The agent proposes; the tool validates before apply.

QY

Query

Read bounds, counts, and selection state so the model can replan instead of guessing the scene is empty.

OK

Validate

Structural checks: duplicates, impossible orientations, support assumptions — fail loud, not silent.

//

Batch

Multi-step operations as one intent: grids, corridors, floors — still a list of typed calls under the hood.

RX

Session

Receipt-style logging of what ran. Useful for debugging the agent, not a public dump of proprietary plans.

Boundary · Northbridge-style honesty

What this page will never publish.

Capability demo ≠ open-sourcing the moat. If it’s measurement, catalog, or bridge internals, it stays private.

Shown here

  • Control-loop idea: NL → MCP tools → editor side effects
  • Layer diagram (agent / server / bridge)
  • Abstract capability verbs (place, transform, query…)
  • Simulated session with sanitized tool names
  • Honest “private prototype” status

Not shown

  • Full tool schemas, ports, or listener internals
  • Asset path catalogs or proprietary name maps
  • Measured pivot / yaw / floor constants
  • Live execute payloads or plan JSON dumps
  • Credentials, endpoints, or editor install paths