Unreal MCP Server — 370+ AI Tools for UE5 Editor Automation via MCP

Bridging an AI coding assistant to a complex game engine often means juggling fragile scripts or burning through token limits just to enumerate available commands. Unreal MCP Server v4 attacks this friction head-on, connecting MCP-compatible agents directly to the Unreal Engine 5.7 Editor.

The server ships 378 editor tools across 54 categories alongside 15 context resources and 14 workflow prompts. It communicates via JSON-RPC 2.0 over Streamable HTTP, turning environments like Claude Code, Cursor, Windsurf, Claude Desktop, VS Code (Copilot), Antigravity, and OpenCode into functional co-developers for level building, gameplay scripting, cinematics, VFX, audio, networking, and animation.

Catalog Mode and AI Context Management

Passing the schemas for 378 tools to an AI model at startup consumes roughly 60,000 tokens. The v4 release defaults to Catalog mode, a progressive tool disclosure system that trims a fresh session's tool-definition cost by roughly 95%.

In Catalog mode, the tools/list endpoint returns four discovery meta-tools and around 30 core tools totaling about 3,000 tokens. Agents discover the remaining tools on demand using search_tools, list_tool_categories, and get_tool_schemas. Every tool remains callable despite not being loaded into the initial context. Developers can switch to the old behavior by setting ToolExposureMode to Full, or instantly manage AI context costs using five tool presets: Full, Scene Building, Gameplay, Minimal, and Catalog.

Batch Scene Building with run_tool_script

Executing batch operations in an engine historically requires multiple round-trips between the agent and the editor. The run_tool_script command eliminates this by running multi-step programs as a single, all-or-nothing editor transaction.

Scripts accept sequential tool calls with result references, such as save_as and $var.field, alongside foreach loops. If any step fails, the entire sequence rolls back automatically. A 24-Actor level blockout becomes a single call and requires a single Ctrl+Z to revert, solving the N-round-trip problem for batch work.

Closed-Loop Blueprint Topology Editing

Reading visual scripting state back to an AI agent normally takes dozens of queries to map pins and connections. The describe_graph function returns full Blueprint topology in one call, including nodes, pins, a deduplicated edge list, and compile status.

This gives the agent the ability to verify its own edits. The 60 Blueprint tools include components, variables, wiring nodes, setting properties, full graph read-back, and cleanup. Developers can trace execution paths, move or delete nodes, manage comments and reroutes, find orphaned nodes, list node types, and query Actor components or component info without leaving the AI prompt context.

Real-Time Modeling and Chaos Destruction

Previous schema stubs for modeling and physics are fully implemented in v4. The five modeling tools—polyextrude, polycut, boolean, uv_unwrap, and remesh—run via GeometryScript and report before and after triangle counts.

Chaos Destruction tools now execute for real. Developers can use chaos_create_geometry_collection and chaos_fracture with Voronoi, cluster, uniform, and planar settings. The chaos_apply_force command applies runtime impulses to generated geometry.

Background Tasks and Structured Agent Feedback

Long-running operations like lighting and navigation builds, fracture calculations, remeshing, PCG execution, fal.ai generation, and automation runs no longer block the editor. These tools return a task ID and a working status after a 3-second grace window. Agents poll get_task_status for progress, abort with cancel_task, and enumerate active jobs with list_tasks. This runs as standard tools, so every MCP client handles long operations without capability negotiation.

Every JSON-returning tool provides structuredContent alongside human-readable text. Structured errors return a machine-readable code, hint, and did_you_mean array. Agents switch on the error code for recovery flows instead of regexing strings, which cuts retry loops.

Engine Integration and Production Safety

The server works against any source control provider Unreal Engine supports, including Perforce, Git LFS, Plastic SCM, and Subversion. AI-powered image generation and text/image-to-3D generation route through fal.ai. An STDIO bridge handles non-HTTP clients like Claude Desktop.

Safety mechanics include bearer-token authentication, an origin allow-list defaulting to localhost-only, scope gating, rate limiting, and dry-run capabilities. The core rejects oversized request bodies with an HTTP 413 error before parsing, enforces tool-call timeouts, and uses exception guards to return structured internal errors instead of crashing the editor. Idle sessions trigger garbage collection, and DNS-rebinding defense validates Host headers.

Automated Testing for the Tool Registry

Quality control relies on MatrixSpec, which generates contract tests across the entire tool registry. New tools are covered the moment they register. During development, this system caught a v3 editor crash and five silent bugs. The export_tool_docs command regenerates the per-tool reference documentation from the live registry, ensuring docs never drift from the actual tool capabilities.

Every mutating operation runs inside its own named MCP editor transaction. Ctrl+Z behaves as expected, and multi-step scripts collapse into one undo step. A redesigned status bar includes a live indicator and one-click access to copy the Claude Code configuration.

Unreal MCP Server fits into production pipelines where artists and developers need autonomous agent execution without sacrificing engine stability or manual revision control.

Related Resources Worth Checking

• NeoStack AI
• LE Encryption
• VibeUE
• AI Engine Assistant - Kibibyte Labs (Blueprint Generator & More!)
• Ultimate Engine CoPilot (formerly Ultimate Blueprint Generator)