Skip to content
Media 1 for listing Unreal MCP Server — 370+ AI Tools for UE5 Editor Automation via MCP
0 comments

Description

Connect Claude Code, Cursor, Windsurf, Claude Desktop, VS Code (Copilot), Antigravity, OpenCode, or any MCP-compatible AI agent directly to the Unreal Engine 5.7 Editor.

The v4 release ships 378 editor tools across 54 categories, plus 15 context resources and 14 workflow prompts via JSON-RPC 2.0 over Streamable HTTP — turning your AI assistant into a true co-developer for level building, gameplay scripting, cinematics, VFX, audio, networking, animation, UI, source control, automated testing, runtime debugging, modeling, material layers, Chaos destruction, and MetaHumans.

🎬 Video Showcase https://www.youtube.com/watch?v=wEuaFqUPp00

🚀 What's New in v4 (vs v3) v4 is the agent-native release — rebuilt around how AI agents actually work. Same workflow, far cheaper context, far safer multi-step editing.

Catalog mode (default) — ~95% less agent context

  • tools/list returns 4 discovery meta-tools + ~30 core tools (~3K tokens) instead of all 378 schemas (~60K tokens). Agents discover the rest on demand via search_tools, list_tool_categories, and get_tool_schemas. Everything stays callable. Set ToolExposureMode = Full for the old behavior.

run_tool_script — multi-step programs as one transaction

  • Sequential tool calls with result references (save_as / "$var.field") and foreach loops, executed in ONE editor transaction (all-or-nothing, rolled back on first failure). A 24-actor level blockout becomes a single call and a single Ctrl+Z — killing the N-round-trip problem for batch work.

Background tasks — long ops no longer block

  • Long-running tools (lighting/nav builds, fracture, remesh, PCG execute, fal.ai generation, automation runs) return {task_id, status: "working"} after a 3s grace window. Poll get_task_status for progress, abort with cancel_task, enumerate with list_tasks. Implemented as tools, so every MCP client works without capability negotiation.

Closed-loop graph editing — agents verify their own edits

  • describe_graph returns full Blueprint topology (nodes + pins + deduplicated edge list + compile status) in ONE call, replacing N×get_node_pins round-trips. Plus get_execution_paths, move_node, delete_nodes, comments, reroutes, find_orphaned_nodes, list_node_types, and list_actor_components / get_component_info.

Modeling & Chaos are now real (were schema stubs in v3)

  • Modeling (5): polyextrude, polycut, boolean, uv_unwrap, remesh — implemented via GeometryScript, with before/after triangle counts.

  • Chaos: chaos_create_geometry_collection and chaos_fracture (voronoi / cluster / uniform / planar) now execute for real; chaos_apply_force applies runtime impulses.

Structured output everywhere

  • Every JSON-returning tool ships structuredContent alongside human-readable text. Structured errors carry a machine-readable {code, hint, did_you_mean[]} — agents switch on code for recovery flows instead of regexing strings, cutting retry loops dramatically.

Universal undo

  • Every mutating tool now runs inside its own named MCP: <tool> editor transaction. Ctrl+Z works as expected; multi-step scripts collapse into one undo step.

Hardened core

  • Request body size limit (HTTP 413 before parsing), tool-call timeouts (structured timeout error), exception guards (a throwing tool returns a structured internal error instead of crashing the editor), idle-session garbage collection, a real BindAddress setting, and DNS-rebinding defense via Host-header validation.

Quality guarded by a registry-wide test matrix

  • MatrixSpec generates contract tests over the ENTIRE registry. New tools are covered the moment they register; it caught a v3 editor crash and 5 silent bugs during development. export_tool_docs regenerates the per-tool reference from the live registry so docs can't drift.

Redesigned status bar

🚀 Key Features

  • 378 Editor Tools across 54 categories — Actors, Materials, Material Layers/Graph, Blueprints, Sequencer, Landscapes, Foliage, Niagara, Audio, MetaSounds, MetaSound Graph, Physics, Chaos, Navigation, PCG, GAS, Enhanced Input, Networking, AI Behavior Trees, AnimGraph, StateTree, Control Rig, CommonUI, AI 3D/Image gen, Source Control, Test Authoring, Runtime Debug, PIE, Modeling, MetaHuman, plus the v4 Meta/Tasks discovery layer.

  • Catalog Mode — progressive tool disclosure cuts a fresh session's tool-definition cost ~95% (~3K vs ~60K tokens) while keeping all 378 tools callable.

  • Tool Presets — Full / Scene Building / Gameplay / Minimal / Catalog. Switch instantly to manage AI context cost.

  • Multi-step tool scripts — run_tool_script runs a sequence (with variables and loops) as ONE all-or-nothing editor transaction.

  • Full Undo Support — every mutating operation is its own named editor transaction; scripts collapse into one undo step.

  • Universal AI Compatibility — Claude Code, Claude Desktop, Cursor, VS Code (Copilot), Windsurf, Antigravity, OpenCode, and any MCP-compatible client. Streamable HTTP + SSE — no vendor lock-in.

  • 15 Context Resources — project data, level hierarchy, selected actors, performance stats, capability detection, undo history, and more.

  • 60 Blueprint Tools — components, variables, wire nodes, set properties, full graph read-back, exec-path tracing, and cleanup.

  • Source Control Aware — works against any provider Unreal supports (Perforce, Git LFS, Plastic, Subversion).

  • Production-grade safety — bearer-token auth, origin allow-list, scope gating, structured errors, cooperative cancellation, dry-run, request size limits, tool-call timeouts, exception guards, session GC, DNS-rebinding defense.

📦 What's Included

  • Full C++ source code with extensible architecture

  • 378 tools across 54 categories

  • 5 tool presets (Full / Scene Building / Gameplay / Minimal / Catalog)

  • 15 read-only context resources

  • 14 reusable workflow prompts

  • AI-powered image generation and text/image-to-3D via fal ai

  • STDIO bridge for non-HTTP clients (Claude Desktop)

  • Live status bar indicator with one-click Copy Claude Code Config

  • Bearer-token auth, origin allow-list, scope gating, rate limiting

  • Registry-generated tool reference (export_tool_docs) and comprehensive documentation

🧩 Tool Categories (54) Core — Actor (16), Editor (7), Asset (6), Level (6), Asset Management (7), Batch (3) Scene Building — Material (5), Material Graph (8), Material Layers (4), Static Mesh (7), Environment (4), Modeling (5) Scripting — Blueprint (60), Python Bridge (1), Search (2) Cinematic — Sequencer (12), Animation (5), Anim Graph (14), Control Rig (2) World Building — Landscape (3), Foliage (4), Spline (7), World Partition (2), Spatial (10), PCG (9) VFX & Audio — Niagara (3), Audio (3), MetaSound (6), MetaSound Graph (8) Simulation — Physics (9), Navigation (3), Chaos (5) Characters — MetaHuman (4) Data — Data Tools (6) UI — Widget/UMG (15), CommonUI (4), UI Image (2) Gameplay — GAS (8), Enhanced Input (6), Game Framework (6), Gameplay Tags (3), State Tree (5), PIE (9) Networking — Networking (5) AI — AI Tools (8), 3D Model (3) DevOps — Source Control (8), Test Authoring (5), Runtime Debug (7) Workflow — Macro (6), Build (7), Engine API (3), Performance (4) Agent (v4) — Meta & Tasks (8)

🔌 Plugin Dependencies (auto-enabled) EditorScriptingUtilities, Niagara, PCG, EnhancedInput, SourceControl, FunctionalTesting, GeometryScripting, Fracture, GameplayAbilities (optional), PythonScriptPlugin (optional)

⚙️ Technical Specs

  • Protocol: MCP spec 2025-06-18, Streamable HTTP

  • Transport: JSON-RPC 2.0 over HTTP (default port 13579) + SSE fallback

  • Engine Support: Unreal Engine 5.7 (Windows, macOS, Linux)

  • Threading: All UE API calls automatically dispatched to the game thread

  • Source Code: Full C++ access included

  • Auth: Bearer-token + origin allow-list (localhost-only by default)

  • Documentation: https://www.strayspark.studio/docs/unreal-mcp-server

Included formats

Tags