中文版点击这里

UE5-UMG-MCP 🤖📄

A Version-Controlled AI-Assisted UMG Workflow

🎬 Click to watch what UMG MCP can do

• Demo Designed A RTS UI
• Demo Recreating the UE5 editor
• Demo Recreating the UE5 editor in UMG editor
• Chat with Gemini 3 to editor the UMG file

---

🛍️ Looking for an easier Setup? → Purchase UMG MCP on Fab Marketplace

If you find manual plugin configuration and MCP environment setup too cumbersome, check out our commercial version on Fab: UMG MCP on Fab Marketplace

Differences between Fab and Open Source versions:

• Out-of-the-Box: The Fab version installs directly via the Unreal Engine launcher, eliminating the need for manual cloning and Python environment setup.
• Context Compression: The Fab version includes advanced context management, automatically filtering redundant MCP history to maximize available tokens for the AI.
• Commercial Focus: While the Open Source version is our protocol testbed, the Fab version is optimized for production business logic. Commercial users receive prioritized support for their specific integration challenges.
• Maintained Parity: We remain committed to open source; all core logic improvements will continue to be synced to this GitHub repository.

---

👉 View Developer Program — contribute to the project and get Fab access for free.

---

🚀 Quick Start

This guide covers the two-step process to install the UmgMcp plugin and connect it to your Gemini CLI.

• Prerequisite: Unreal Engine 5.6 or newer.

1. Install the Plugin

1. Navigate to your project's Plugins folder: YourProject/Plugins/ (create it if it doesn't exist).

2. Clone the repository directly into this directory:

   git clone https://github.com/winyunq/UnrealMotionGraphicsMCP.git UmgMcp

3. Restart the Unreal Editor. This allows the engine to detect and compile the new plugin.

2. Connect the Gemini CLI

Tell Gemini how to find and launch the MCP server.

1. Edit your settings.json file (usually located at C:\Users\YourUsername\.gemini\).

2. Add the tool definition to the mcpServers object.

   "mcpServers": {
     "UmgMcp": {
       "command": "uv",
       "args": [
         "run",
         "--directory",
         "D:\\Path\\To\\YourUnrealProject\\Plugins\\UmgMcp\\Resources\\Python",
         "UmgMcpServer.py"
       ]
     }
   }

   IMPORTANT: You must replace the path with the correct absolute path to the Resources/Python folder from the cloned repository on your machine.

That's it! When you start the Gemini CLI, it will automatically launch the MCP server in the background.

Testing the Connection

After restarting your Gemini CLI and opening your Unreal project, you can test the connection by calling any tool function:

  cd Resources/Python/APITest
  python UE5_Editor_Imitation.py

Python Environment (Optional)

The plugin's Python environment is managed by uv. In most cases, it should work automatically. If you encounter issues related to Python dependencies (e.g., uv command not found or module import errors), you can manually set up the environment:

1. Navigate to the directory: cd YourUnrealProject/Plugins/UmgMcp/Resources/Python
2. Run the setup:

   uv venv
   .\.venv\Scripts\activate
   uv pip install -e .

---

🧪 Experimental: Gemini CLI Skill Support

We are experimenting with the Gemini CLI Skill system as an alternative to the standard MCP approach. The Skill architecture allows the Python tools to be loaded directly by the CLI runtime, potentially optimizing context usage by dynamically enabling/disabling tools via prompts.json and avoiding the overhead of managing a separate MCP server process.

Note: The MCP server (configured above) is still the stable and recommended way to use this plugin. Use Skill mode if you want to test the latest integration capabilities.

Configuration (Skill Mode)

To enable Skill mode, add the following to your settings.json (replacing <YOUR_PROJECT_PATH>):

  "skills": {
    "unreal_umg": {
      "path": "<YOUR_PROJECT_PATH>/Plugins/UmgMcp/Resources/Python/UmgMcpSkills.py",
      "type": "local",
      "description": "Direct control of Unreal Engine UMG via Python Skills. Auto-loads tools from prompts.json."
    }
  },

---

This project provides a powerful, command-line driven workflow for managing Unreal Engine's UMG UI assets. By treating human-readable .json files as the sole Source of Truth, it fundamentally solves the challenge of versioning binary .uasset files in Git.

Inspired by tools like blender-mcp, this system allows developers, UI designers, and AI assistants to interact with UMG assets programmatically, enabling true Git collaboration, automated UI generation, and iteration.

---

📋 Prompt Manager

A visual web tool for configuring system instructions, tool descriptions, and user prompt templates.

Features

1. System Instruction Editor: Modify the global instructions for the AI context.
2. Tool Management:
   • Enable/Disable: Toggle specific MCP tools on or off. Disabled tools are not registered with the MCP server, effectively compressing the context window to prevent AI distraction.
   • Edit Descriptions: Customize tool descriptions (prompts) to better suit your workflow.
3. User Templates (Prompts): Add reusable prompt templates for quick access by the MCP client.

How to Run

Execute the following command in your Python environment:

python Resources/Python/PromptManager/server.py

The browser will automatically open http://localhost:8085.

Usage Tips

Prompts are crucial for AI tool effectiveness. Use the Prompt Manager to tailor the AI's behavior:

• One-Click Deployment Mode: If you want the AI to focus solely on generating UI from design, disable all tools except apply_layout and export_umg_to_json.
• Tutor Mode: If you want the AI to guide you without making changes, keep only read-only tools (e.g., get_widget_tree, get_widget_schema).
• Context Optimization: For models with smaller context windows, disable tools you aren't currently using to improve speed and accuracy.

Contributions of effective prompt configurations are welcome!

---

🏗️ Current Technical Architecture

The system now primarily relies on the UE5_UMG_MCP plugin for communication between external clients (like this CLI) and the Unreal Engine Editor.

Architecture Diagram:

flowchart LR
    subgraph "Local Execution Environment"
        CLI["Gemini CLI"] --"StdIO (JSON-RPC)"--> PY["Python (UmgMcpServer.py)"]
    end

    subgraph "Unreal Engine 5"
        PY --"TCP Socket (JSON)"--> TCP["UmgMcpBridge (C++)"]
        TCP --> API["Unreal API / UMG"]
    end

Loading

---

⚖️ AI Authorship & Disclaimer

This project has been developed with significant assistance from Gemini, an AI. As such:

• Experimental Nature: This is an experimental project. Its reliability is not guaranteed.
• Commercial Use: Commercial use is not recommended without thorough independent validation and understanding of its limitations.
• Disclaimer: Use at your own risk. The developers and AI are not responsible for any consequences arising from its use.

---

API Status

Category	API Name	Status
Context & Attention	get_target_umg_asset	✅
	set_target_umg_asset	✅
	get_last_edited_umg_asset	✅
	get_recently_edited_umg_assets	✅
Sensing & Querying	get_widget_tree	✅
	query_widget_properties	✅
	get_creatable_widget_types	✅
	get_widget_schema	✅
	get_layout_data	✅
	check_widget_overlap	✅
Actions & Modifications	create_widget	✅
	delete_widget	✅
	set_widget_properties	✅
	reparent_widget	✅
	save_asset	✅
File Transformation	export_umg_to_json	✅
	apply_json_to_umg	✅
	apply_layout	✅

UMG Blueprint API Status (New)

Category	API Name	Status	Description
Context & Attention	set_edit_function	✅	Set the current edit context (Function/Event). Supports auto-creating Custom Events.
	set_cursor_node	✅	Explicitly set the "Cursor" node (Program Counter).
Sensing & Querying	get_function_nodes	✅	Get nodes in Current Context Scope (Filtered to connected graph to avoid global noise).
	get_variables	✅	Get list of member variables.
	search_function_library	✅	Search callable libraries (C++/BP). Supports Fuzzy Search.
Actions & Modifications	add_step(name)	✅	Core: Add executable node by Name (e.g. "PrintString"). Auto-Wiring & Auto-Layout supported.
	prepare_value(name)	✅	Add Data Node by Name (e.g. "MakeLiteralString", "GetVariable").
	connect_data_to_pin	✅	Connect pins precisely (Supports NodeID:PinName format).
	add_variable	✅	Add new member variable.
	delete_variable	✅	Delete member variable.
	delete_node	✅	Delete specific node.
	compile_blueprint	✅	Compile and apply changes.

UMG Sequencer API Status

Command	Status	Description
animation_target	✅	Set/focus the current animation (alias of set_animation_scope, auto-creates when missing).
widget_target	✅	Set/focus the current widget (alias of set_widget_scope).
animation_overview	✅	Returns keyframe counts, track counts, key times, and changed properties.
animation_widget_properties	✅	Timeline view: per-widget property changes (ignores unanimated properties).
animation_time_properties	✅	Time-slice view: property values at specific times (multi-time supported).
animation_append_widget_tracks	✅	Append/overwrite keys per widget+property (union only, no implicit deletion).
animation_append_time_slice	✅	Append a diff-style time slice for multiple widgets at a given time.
animation_delete_widget_keys	✅	Scoped delete for widget+property at specific times (confirm_delete=true required per Issue 15 safety policy).
create_animation	✅	Create or focus an animation with auto naming.
set_property_keys	✅	Low-level track write helper (supports float/color/vector2D).

Notes:

• animation_target/widget_target reuse the current UMG target asset; names are auto-corrected (no “animal” typo) and auto-create when missing.
• Write paths are union/overwrite only—no implicit deletion. Use animation_delete_widget_keys with confirm_delete=true for scoped removals.
• Responses now include counts/timeline context so every sequencer MCP returns actionable data.

UMG Material API Status (New: The 5 Core Pillars Strategy)

Category	API Name	Status	Description
P0: Context	material_set_target	✅	Anchor: Specify target asset (path or parent). Required for stateful editing.
P1: Def & Query	material_define_variable	✅	Define external interface parameters (Def, not wire). Supports Scalar, Vector, Texture.
P2: Symbol Place	material_add_node	✅	Drag Symbol: Place a symbol from lib into graph and assign a unique instance handle.
	material_get_graph	✅	Query existence and state of node instances in the graph.
P3: Connectivity	material_connect_nodes	✅	Natural Connection: Establish node-to-node functional flow (A -> B).
	material_connect_pins	✅	Surgical Wiring: Manually connect specific pins for complex topologies.
P4: Lib Search	material_search_library	🚧 Planned	Search for available Material Expressions (symbols) and Functions.
P5: Tactical Detail	material_set_hlsl_node_io	✅	Tactical Code: Inject HLSL into Custom nodes and sync IO pins via JSON mapping.
	material_set_node_properties	✅	Property Editing: Set internal properties for regular nodes (e.g. Constant Value, Texture).
Lifecycle	material_compile_asset	✅	Submit changes and analyze HLSL compilation errors.
Maintenance	material_delete	✅	Delete node instances or clean up logic by unique handle.
	material_get_pins	✅	Introspect pins for a specific node handle.

UMG HLSL MCP API Status (New: Text-Edit Loop for UMG)

Command	Status	Description
hlsl_set_target	✅	Lock/create HLSL target material. Validates UI-domain + single Custom node topology; can request confirmation before overwrite.
hlsl_get	✅	Read current HLSL code and structured input parameters from the single Custom node.
hlsl_set	✅	Incremental update of HLSL and/or parameters. Deletion is explicit (delete: true) to avoid accidental destructive edits.
hlsl_compile	✅	Compile current HLSL target and return concise diagnostics for AI post-processing.

HLSL Protocol Contract (UMG-Optimized)

• Material is treated as a single HLSL program.
• Backend assumes HLSL returns float4.
• Output auto-wiring is fixed: .rgb -> FinalColor, .a -> Opacity.
• Input parameters are returned as structured descriptors (name, kind, source_handle) for learning/replay by AI agents.
• hlsl_set is safety-biased: writing is easy, deletion must be explicit.

UMG Style & Theming API Status (New)

Category	API Name	Status	Description
Styling	set_widget_style	🚧 Planned	Set detailed style (e.g. FButtonStyle) for a specific widget.
Theming	apply_global_theme	🚧 Planned	Batch apply styles and fonts across multiple widgets based on a theme config.
Assets	style_create_asset	🚧 Planned	Create a standalone Slate Widget Style asset.

---

Developer Program

We notice there are many forks but few PRs — here's your invitation to change that.

❓ What is the Developer Program & Why does it exist?

We are living in the age of AI. Everyone now has the ability to build and contribute to projects with AI assistance. UMG MCP is sincerely provided free of charge for everyone to learn and use — and this should include the Fab version too.

By joining the Developer Program and meeting the contribution criteria, you will gain access to the private repository during your active development period.

🎁 Why is the reward a Fab version license?

The only reason the Fab version is paid is an economic reality: if it were free, the average social return on labor would drop dramatically. Charging for it is the only mechanism to sustain development. The Fab version of UMG MCP should be free — but making it truly free would only force us to work harder for less. Therefore, rewarding contributors with a Fab license serves two purposes: it gives you access to the more polished version, and it gives you the ability to maintain the real, production-grade project.

🛠️ How to join the UMG MCP Developer Program?

This is admittedly a bit of a paradox — UMG MCP is designed to serve non-programmers, and programmers may not need it as much. Regardless, here are the paths to contribute:

Path 1 — Video Content: Create a video specifically about the UMG MCP Fab version. Reach a meaningful audience and we'll count it.

Path 2 — Feature Development: Our design philosophy is simple: if your tool gets accepted, you're in. Develop a feature, submit a PR, and if we merge it, you qualify.

Path 3 — Prompt Engineering: Write system prompts that help the AI more accurately identify and invoke the correct tools — even when all tools are enabled simultaneously. Precision matters here.

Path 4 — Purchase the Fab version: If you've purchased it, you already have the right to access this project. Simple as that.

To apply: Send your GitHub profile to winyunq@gmail.com