VibeUE

Unreal 5.8 ships its own AI toolsets (Blueprints, materials, actors, assets, meshes, data tables, …). VibeUEcomplementsthem — it focuses on the domains and depth the engine doesn't cover:

- Terrain & world— Landscape sculpting/heightmaps/splines, landscape auto-materials + RVT, Foliage, procedural FPSMap Blockout, andreal-world terrain(heightmaps + water from GPS).
- Audio— MetaSound and SoundCue graph authoring.
- Animation assets— AnimSequence keyframe editing, AnimMontage authoring, AnimBP state machines, Skeleton bone/socket/retarget/blend-profile editing.
- FX depth— Niagara emitter color/curve authoring andCustom-HLSL scratch-padmodules.
- UI— UMG widgets with MVVM bindings, animation authoring, and preview/PIE validation.
- Higher-order Blueprint authoring— timelines, event dispatchers, delegates, custom-event pins, comment boxes, and a batchbuild_graphbuilder.
- AI behavior authoring— Behavior Tree node/decorator/service editing by structural path with blackboard key binding, JSON tree round-trips, and validate/repair; full Blackboard asset + key CRUD. Writes go through the editor EdGraph with PIE- and open-editor-aware refusal guards.
- Editor safety—TransactionServicewraps the editor's transaction buffer (undo / redo / checkpoints) so an agent can group and roll back its own edits — the engine's toolsets expose none.
- ⚡ Performance & profiling— VibeUE's standout: see the dedicated section below.
- Python-first access— run anyunreal.Python in the editor and introspect the whole API.
- Web research— search / fetch / geocode for in-context research and terrain workflows.

Unreal's native AI toolsets havezeroperformance tooling— they can start PIE/Simulate but can't measure anything. VibeUE'sPerformanceServicefills that gap so an agent can actuallydiagnose and fixframe rate:

- frame_timing()— Game/Render/GPU/RHI thread split + aCPU-vs-GPU-bound verdictand a concrete next-step hint.Run this first— optimising the GPU does nothing on a CPU-bound frame.
- Unreal Insights capture—start_trace/stop_trace/get_trace_status, withbookmarkandregion_start/region_endmarkers.
- analyse()— reads the traceandlog back and returns a perf summary (frame stats, worst frames, hitches, notable log lines).
- Trace-attachedstart_standalone— profile a representative standalone build, not just the editor viewport.

import unreal print(unreal.PerformanceService.frame_timing()) # CPU vs GPU bound — diagnose FIRST unreal.PerformanceService.start_trace("cap", "") # Insights trace # … reproduce the workload (ideally under PIE / standalone) … unreal.PerformanceService.stop_trace() print(unreal.PerformanceService.analyse("both", ""))

Step 1 — Set up Unreal's native MCP (do this FIRST)

cd /path/to/YourProject/Plugins git clone https://github.com/kevinpbuckley/VibeUE.git

Plugins/VibeUE/BuildAndLaunchGame.ps1 # builds + launches the editor Plugins/VibeUE/BuildAndLaunchGame.ps1 -StrictRebuild # full recompile (warnings-as-errors)

On Linux or macOS, use the platform-detecting shell script:

Plugins/VibeUE/BuildAndLaunchGame.sh --engine /path/to/UE5 Plugins/VibeUE/BuildAndLaunchGame.sh --engine /path/to/UE5 --strict-rebuild

ThenEdit → Plugins→ enableVibeUEand restart. Its services, tools, and skills now register onto the same endpoint, alongside the engine's own.

Two console commands from the editor (open the console with``):

1. Write the MCP server config(.mcp.jsonat the project root):

ModelContextProtocol.GenerateClientConfig ClaudeCode

(supportsClaudeCode,Cursor,VSCode,Gemini,Codex, orAll.)

This writes the guide to the correct file for your agent —CLAUDE.md(Claude Code),GEMINI.md(Gemini),AGENTS.md(Codex / Hermes / Cursor), or.github/copilot-instructions.md(Copilot) — or passAllto write CLAUDE.md + GEMINI.md + AGENTS.md at once. It resolves the plugin location automatically, so it works whether VibeUE was installed fromFABorGit. The guide goes in a managed block, so re-run any time to refresh without disturbing your own notes. Passimportto link the guide with a one-line@importinstead of copying it (Claude Code / Gemini only — other agents don't resolve imports, so they always get a copy).

The guide teaches: discover before you call (discover_python_class), batch withexecute_python_code, load skills viaListSkills/GetSkills, and when to reach fordeep_research/terrain_data.

Skills are lazy-loaded domain knowledge (workflows, gotchas, property formats) served by Unreal's nativeAgentSkillToolset:

This is a web browser that enables your coding agent, such as Claude Code, to visit websites on your behalf and assist you in identifying bugs or creating UI test cases.

An MCP server for Unity, enabling AI assistants to interact with projects in real-time, access scene data, and execute code.

Manipulate Adventure Game Studio (AGS) compiled room (.crm) files to enable AI-powered game development.

Godot Engine integration: 17 tools for game development

An MCP server for AI coding assistants to control, inspect, and modify Bevy applications using the Bevy Remote Protocol (BRP).

Control, inspect, and mutate Bevy applications with AI coding assistants via the Bevy Remote Protocol (BRP).

A server for Ebitengine games that provides debugging and recording tools by capturing game state.

Fennara MCP connects AI agents like Codex, Cursor, Claude Code, and Claude Desktop to Godot-aware tools for real Godot projects. It focuses on feedback from Godot: GDScript diagnostics, scene validation, runtime errors, scene inspection, node properties, screenshots, SemanticSearch, and patch-and-rerun workflows.

an open-source MCP server that unifies Roblox Studio, Unity, Unreal Engine, and Blender into a single AI control plane for game development workflows.

7 tools for 3D game development — character viewers, level editors, physics games, particle effects, 3D inventories with SceneView. 156 tests.