Skip to content

AI Integration

Driving UE from LLMs and AI tools. Two distinct categories:

Category Examples Section
In-editor AI tooling — LLM controls the UE editor over a socket UnrealClaudeMCP, Unreal AI Connection This page
In-game AI — Behavior Trees, EQS, Mass AI (not covered here — separate field) UE docs

Pages

What problem are we solving?

UE's Python API has known dead-ends (e.g. EditorUtilityWidgetBlueprint.WidgetTree not exposed). LLM-driven workflows need:

  1. A stable, JSON-RPC-style protocol the LLM can call
  2. Native C++ handlers that bypass Python reflection limits
  3. Comprehensive tool surface (asset registry, level editing, BP introspection, Python escape hatch)
  4. A local TCP path (no cloud, no auth dance, instant)

UnrealClaudeMCP solves this.

MCP (Model Context Protocol)

Anthropic's vendor-neutral protocol for LLM tool use. Each tool exposes a name, JSON schema, and a callable handler. Clients (Claude Code, Cursor, Codex CLI, …) speak MCP over stdio to a bridge process; the bridge proxies to UE via local TCP.

graph LR
  A[MCP client e.g. Claude Code] -->|stdio MCP| B[Python bridge]
  B -->|TCP 127.0.0.1:18888| C[UE plugin]
  C -->|native C++| D[UE editor]

Quick start

  1. Install UnrealClaudeMCP plugin into your UE project (Plugins/)
  2. Build editor — handlers register on startup
  3. Configure your MCP client to point at the bridge script
  4. Test with examples/smoke_test.py

See the per-tool pages for full reference.

Workflows enabled

  • "Spawn a cube at origin, focus the camera on it, take a screenshot"
  • "Inspect this Editor Utility Widget and propose a layout fix"
  • "Audit all Blueprint compile errors"
  • "Bulk rename all assets in /Game/Items to PascalCase"
  • "Compile a mod PAK"
  • "Find unused assets and list their disk footprint"

All of these are tool calls the LLM emits, not "type in the editor for me" pixel-clicking.