ue5-mcp
A field manual for AI agents driving Unreal Engine 5 through MCP.
When an AI agent connects to the UE5 editor through an MCP server, it gains access to tools for manipulating Blueprints, materials, Niagara particles, MetaSound audio, meshes, widgets, levels — anything UE's reflection surface exposes. The tools are powerful but UE is full of undocumented quirks: APIs that silently no-op, crash patterns that only surface at runtime, render paths that compile cleanly but produce nothing visible, async operations that don't block, and reflection rules that bite anyone who treats the engine like a typical scripting target.
This skill gives your AI that knowledge upfront so it doesn't have to rediscover it every session.
The skill is server-agnostic. It works with any MCP server that exposes UE5 functionality — including Epic's official ModelContextProtocol plugin (experimental, ships with UE 5.8), and any third-party / custom server. It documents engine-level behaviour, not server-level features. Server-specific tool names and shapes vary; use tools/list against your own server to discover what's actually available.
What's inside
- UE5 reflection gotchas — PascalCase vs snake_case silent no-ops, Blueprint class path
_C suffix, async asset operations, PostEditChangeProperty requirement, three-step graph mutation
- Crash patterns — referenced-mesh deletion, rapid actor ops, MetaSound scalar-to-audio-pin, save-before-PIE for Niagara/MetaSound
- Subsystem-specific knowledge — Lumen mobility, Blueprint instance override staleness, Niagara create-from-empty /
script_usage / dynamic input limitations, MetaSound pin exactness, material emissive bloom threshold + translucent shading, UMG CreateWidget owner context, in-world widget components, AudioComponent reflection name
- Python ↔ MCP data channel — Actor Tags workaround for Python
print() output that goes to the UE log instead of back to MCP
- MCP transport requirements — Accept header, image content blocks, cancellation semantics, sessions
- Patterns for MCP server authors and clients — schema-in-error, output-schema authority, continuation tokens,
FScopedTransaction, verify-after-mutate, recursion caps
- UE 5.7 vs 5.8 differences — engine-level deltas that affect any agent
- Asset structure quick reference — what fields appear in Blueprint, Niagara, Material, Level, Sequencer, and Widget Blueprint dumps
Installation
Claude Code
git clone https://github.com/ibrews/ue5-mcp.git ~/.claude/skills/ue5-mcp
The skill auto-triggers when UE5 MCP tools are detected, or when the user mentions Unreal Engine, UE5, Blueprints, Niagara, MetaSound, materials, or any UE editor automation workflow.
Cowork (manual upload)
- Download this repo as a ZIP
- In Claude desktop: Customize → Skills → "+" → Upload a skill → select the zip.
As a plugin (auto-updates from GitHub)
/plugin marketplace add ibrews/ue5-mcp
/plugin install ue5-mcp@ibrews-ue5-mcp
Works in Claude Code and Cowork. Updates automatically when changes push to this repo.
Other AI agents
The SKILL.md file uses Claude's skill format (YAML frontmatter + markdown), but the content is universal. Feed it into your agent's system prompt or knowledge base. Strip the frontmatter if your framework doesn't understand it — the knowledge sections stand on their own.
Requirements
- Unreal Engine 5.x (this skill covers 5.7 and 5.8 specifically; most content applies to earlier and later versions)
- An MCP server exposing UE5 functionality. Options:
- Epic's official
ModelContextProtocol plugin — experimental, ships with UE 5.8, requires a source build of the engine.
- Third-party / custom MCP servers — the community ecosystem is growing; any MCP server that talks to a running UE5 editor works.
- An AI agent that speaks MCP (Claude Desktop, Cursor, VS Code's MCP UI, Epic's EDA panel, OpenClaw, or any MCP-compatible agent).
- Optional: Pixel Streaming plugin (for visual verification and PIE input).
- Optional: Python Editor Script Plugin (for advanced scripting, required for the Python data-channel pattern).
How it works
The skill is a single SKILL.md file. When loaded, it gives your AI agent context about:
- What will silently fail — APIs that compile clean but no-op, produce no results, or corrupt assets without complaint.
- What will crash the editor — actions that trigger asserts and revert your unsaved work.
- What actually works — call sequences with exact ordering, parameter formats, and the verification steps that matter.
The knowledge is organised by topic (reflection rules, stability hazards, subsystems, transport, patterns) so the agent can quickly find relevant guidance for whatever it's working on.
Things to try
