VibeUE — AI-Powered Unreal Engine Development

🧩 MCP Expansion + AI Editor Toolset for Unreal Engine 5.8+

https://www.vibeue.com/

VibeUE is the MCP Expansion + AI Editor Toolset for Unreal Engine 5.8+. Unreal 5.8 added a built-in MCP server and AI toolsets; VibeUE is an MCP Expansion that plugs straight into them and adds a deep AI Editor Toolset — a library of editor capabilities — Blueprints, materials, landscape, foliage, animation, Niagara, UMG, audio, StateTree, Behavior Trees & Blackboards, gameplay tags, input, UVs, performance/profiling, and more — registered into the engine's own ToolsetRegistry and ModelContextProtocol server, plus rich domain skills served through Unreal's native AgentSkill system. Any MCP-capable agent (Claude Code, Cursor, Copilot, …) drives your editor through Unreal's standard MCP endpoint.

⚠️ VibeUE requires Unreal's native MCP to be set up first — enable the Unreal MCP plugin (which auto-enables Toolset Registry) and the Editor Tools plugin, then start the MCP server. Follow Epic's guide: Unreal MCP in the Unreal Editor. VibeUE then expands that endpoint — no separate server and no in-editor chat. A free API key (set in Editor Preferences → Plugins → VibeUE) unlocks the real-world terrain tools; everything else works without one.

---

✨ What VibeUE adds

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

• Terrain & world — Landscape sculpting/heightmaps/splines, landscape auto-materials + RVT, Foliage, procedural FPS Map Blockout, and real-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 and Custom-HLSL scratch-pad modules.
• 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 batch build_graph builder.
• 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 — TransactionService wraps 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.
• Verified workflows — WorkflowService adds an authoritative engine/build manifest, durable task journals with affected-asset provenance, asynchronous PIE scenarios (compile → input → assertion → capture → guaranteed teardown), and dry-run-first resumable bulk maintenance.
• ⚡ Performance & profiling — VibeUE's standout: see the dedicated section below.
• Python-first access — run any unreal.* Python in the editor and introspect the whole API.
• Web research — search / fetch / geocode for in-context research and terrain workflows.

It deliberately does not duplicate the engine's general tools (basic asset/actor/blueprint/material CRUD, screenshots, logs, PIE) — agents use Unreal's native toolsets for those.

---

⚡ Performance & Profiling (flagship)

Unreal's native AI toolsets have zero performance tooling — they can start PIE/Simulate but can't measure anything. VibeUE's PerformanceService fills that gap so an agent can actually diagnose and fix frame rate:

• frame_timing() — Game/Render/GPU/RHI thread split + a CPU-vs-GPU-bound verdict and 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, with bookmark and region_start / region_end markers.
• analyse() — reads the trace and log back and returns a perf summary (frame stats, worst frames, hitches, notable log lines).
• Trace-attached start_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", ""))

Pair with the profiling and frame-rate skills for the full CPU/GPU drill-down.

---

🏗️ Architecture

VibeUE plugs into three native UE 5.8+ systems:

1. Toolsets (ToolsetRegistry) — VibeUE's services register as UToolsetDefinitions, so their methods become AICallable tools on the MCP endpoint. They're also BlueprintCallable, so the same methods are callable from Python as unreal.<Name>Service.<method>().
2. MCP server (ModelContextProtocol) — a small set of VibeUE utility tools are registered directly on the endpoint: execute_python_code, discover_python_module/_class/_function, list_python_subsystems, deep_research, terrain_data.
3. Skills (AgentSkillToolset) — ~36 markdown skill packs register as native UAgentSkills, discoverable via ListSkills and loaded lazily via GetSkills, alongside the engine's own skills.

Efficient usage (for agents): execute_python_code is the workhorse — it batches a whole multi-step task into one round-trip and reaches every VibeUE service plus the full unreal.* API. Use call_tool only for skills and the few engine