UE-MCP

Unreal Engine Model Context Protocol Server - gives AI assistants deep read/write access to the Unreal Editor through 26 category tools covering 1931+ actions, plus a YAML flow engine for multi-step workflows and an npm plugin system for extending the surface.

That total is 1101+ actions of ue-mcp's own plus Epic's entire native AI Toolset Registry: on UE 5.8+ its 830 official Unreal tools are called in-process and surfaced as epic_* actions in the matching category, Sequencer in animation, PCG in pcg, static meshes in asset.

flowchart LR
    AI[AI Assistant] -->|stdio| MCP[MCP Server<br/>TypeScript / Node.js]
    MCP -->|WebSocket<br/>JSON-RPC| Plugin[C++ Bridge Plugin<br/>inside Unreal Editor]
    Plugin -->|UE API| Editor[Editor Subsystems]
    MCP -->|direct fs| FS[Config INI<br/>C++ Headers<br/>Asset Listing]

Blueprints, materials, levels, actors, animation, VFX, landscape, PCG, foliage, audio, UI, physics, navigation, AI, StateTree, GAS, networking, sequencer, build pipeline - all programmable through natural language.

Filesystem reads (INI config, C++ headers, asset listing) work with the editor closed, so an agent can explore the project before anything is running.

Quick Start

npx ue-mcp init

The interactive setup will:

1. Find your .uproject (auto-detects in current directory)

2. Let you choose which tool categories to enable

3. Deploy the C++ bridge plugin to your project

4. Enable required UE plugins (Niagara, PCG, GAS, etc.)

5. Detect and configure your MCP client (Claude Code, Claude Desktop, Cursor, Codex)

Restart the editor once after setup to load the bridge plugin. To update later: npx ue-mcp update

Then talk to your AI in plain English:

"Check the UE-MCP connection and tell me what's in the current level."
"Place a point light 300 units above the player start, intensity 5000, warm white."
"Create an emissive material that pulses, and apply it to every cube in the level."
"Build the project in Development and report any compile errors."

If it doesn't connect

npx ue-mcp doctor

Documentation lives at https://ue-mcp.com/docs/, one page per file in docs/, published automatically on release. Link there rather than to a file in this repository.

Most first-run failures are one of three things: the editor hasn't been restarted since init, the bridge plugin failed to compile (check the editor's Output Log for UE_MCP_Bridge), or the editor is open on a different project than the one configured. See Troubleshooting.

Manual Configuration

If you prefer to configure manually, add to your MCP client config:

{
  "mcpServers": {
    "ue-mcp": {
      "command": "npx",
      "args": ["ue-mcp", "C:/path/to/MyGame.uproject"]
    }
  }
}

Related MCP server: Monolith

See It Work

UE-MCP ships a built-in 19-step demo that builds a complete procedural scene - landscape, PCG scatter, Niagara VFX, materials, lighting, and a playable blueprint - entirely through the bridge:

demo(action="step", stepIndex=1)   ... through 19
demo(action="cleanup")             restores the level

Nothing is pre-authored. Every asset in the scene is created by the agent at runtime. Full walkthrough: Neon Shrine Demo.

What Can It Do?

Full parameter-level reference: Tool Reference.

Flows

Any sequence of actions can be captured as a YAML flow in ue-mcp.yml next to your .uproject, then run as one call. Config is hot-reloaded, so you edit and re-run without restarting anything.

ue-mcp:
  version: 1

flows:
  build_and_check:
    description: Build the project and verify the editor is connected
    steps:
      1:
        task: project.build
        options:
          configuration: Development
      2:
        task: project.get_status

flow(action="run", flowName="build_and_check")

Every one of the 1931+ actions is also a flow task. Flows support step references, retries, rollback, custom tasks in your own .js/.ts, and shell steps. See Flows.

Plugins

UE-MCP is extensible through npm. A plugin can inject new actions into existing categories, or provision an entirely new category backed by its own C++ handlers.

npm install ue-mcp-<plugin>

Injected actions appear inside the pcg/landscape/etc. tools the agent is already using, so there is nothing extra for it to discover. See Plugins for the consumer view and Plugin Authoring for the contract.

Agent Skills

The package ships skills that teach agents the non-obvious parts of driving the editor, and they load automatically in clients that support them:

Safety

An agent with write access to your project is a real risk, so the bridge is built with brakes:

• Write guards - configurable rules can block or require confirmation on mutating methods before they reach the editor.

• Source control guard - writes to checked-in assets can be gated on source control state instead of silently overwriting.

• Flow rollback - handlers report created/existed/updated and flows can snapshot git state before a run, so a failed multi-step flow doesn't leave debris.

• Idempotency by convention - re-running an action that already applied is a no-op rather than a duplicate.

Details: Guards and Handler Conventions.

Requirements

Editor process control (editor(start_editor) / stop_editor / restart_editor) works on Windows, macOS and Linux. Set UE_EDITOR_PATH if the engine is installed somewhere non-standard.

Documentation

https://ue-mcp.com/docs

• Getting Started - Zero-knowledge walkthrough, install to first actor placed

• Tool Reference - Every tool, action, and parameter

• Architecture - How the server, bridge, and editor fit together

• Flows - Multi-step YAML workflows, custom tasks, rollback

• Journal - The per-project record of what a session did, produced, and how it ended

• Skill Packs - Written workflows an agent reads, and checking that the calls they teach still exist

• Plugins - Extending the surface through npm

• Configuration - ue-mcp.yml and MCP client config

• Troubleshooting - Connection, build, and asset path issues

• Development - Building from source, testing, contributing

Sponsor

Thanks for using UE-MCP. It's free and MIT licensed. If it helps your work, sponsorship is appreciated.

Contributing

Issues and pull requests welcome. Start with Development.

Install git-lfs before cloning - the bundled test project stores .uasset / .umap via LFS, and a plain git clone leaves them as pointer files.

If an AI agent had to fall back to execute_python during your session, it will offer to submit structured feedback automatically. That is how we prioritize which native handlers to add next.

License

UE-MCP is licensed under the MIT License. Use it freely in personal, educational, and commercial projects. See LICENSE.

Contributions are accepted under the same MIT License (inbound contributions are licensed under the project's outbound license).