Install EngineWire in your Unreal project

Before you start

EngineWire is the website for the open-source Unreal_mcp project: a TypeScript MCP server (npm unreal-engine-mcp-server) and a native C++ editor plugin called McpAutomationBridge. The plugin does the work inside the editor; your AI client reaches it over HTTP or stdio. You need:

1. Add the plugin

1. Open the Releases page, pick the newest v0.6 pre-release and download McpAutomationBridge-plugin-<version>.zip. The archive holds source only.

2. The zip contains plugins/McpAutomationBridge/. Copy that McpAutomationBridge folder into your project's Plugins folder:

MyGame/
├── MyGame.uproject
└── Plugins/
    └── McpAutomationBridge/
        ├── McpAutomationBridge.uplugin
        └── Source/

Prefer to track the dev branch? Clone the repository and reference its plugins folder from your .uproject, so a git pull updates the plugin without copying:

{
  "AdditionalPluginDirectories": [
    "C:/Path/To/Unreal_mcp/plugins"
  ]
}

2. Open the project and build

Open the .uproject. When Unreal asks whether to rebuild the missing modules, answer Yes. The plugin is large, so the first build can take several minutes.

Checkpoint: the status bar in the bottom-right of the level editor shows MCP off. That means the plugin loaded and its built-in HTTP server is simply switched off for now. No indicator at all means the plugin did not load.

If you see "Engine modules cannot be compiled at runtime. Please build through your IDE", generate project files, build the Editor target once in Visual Studio, Rider or Xcode, then reopen the project. More fixes are in troubleshooting.

3. Choose a route and connect a client

Both routes expose the same single MCP tool, unreal. Give each client one route, not both, or the tool shows up twice.

Route A: native HTTP

1. In Edit › Project Settings › Plugins › MCP Automation Bridge, under Native MCP, tick Enable Native MCP Server (port 3000 by default). Restart the editor. The status bar now reads MCP :3000 (0): the port, then the number of connected clients.

2. Read the capability token the plugin generated for your project. It is 64 hexadecimal characters. Treat it like a password.

cat /path/to/MyGame/Saved/MCP/capability-token

Get-Content "C:\Path\To\MyGame\Saved\MCP\capability-token"

Then add the server to your client. For Claude Code, one command does it:

claude mcp add --transport http unreal-engine http://127.0.0.1:3000/mcp --header "X-MCP-Capability-Token: <token>"

Cursor and VS Code use a JSON file instead; the exact files are on the Cursor and VS Code pages. When the client connects, the count in the status bar goes up to MCP :3000 (1).

Route B: stdio through Node.js

Add this to your client's MCP configuration. For Claude Desktop the file is %APPDATA%\Claude\claude_desktop_config.json on Windows or ~/Library/Application Support/Claude/claude_desktop_config.json on macOS:

{
  "mcpServers": {
    "unreal-engine": {
      "command": "npx",
      "args": ["-y", "unreal-engine-mcp-server@beta"],
      "env": {
        "UE_PROJECT_PATH": "C:/Path/To/MyGame"
      }
    }
  }
}

UE_PROJECT_PATH is your project folder or its .uproject. The server uses it to find the capability token and the plugin's port, so nothing else needs setting. Restart the client after editing its config. The stdio server starts even when the editor is closed; calls answer NOT_CONNECTED until the editor is running.

4. Try it

With the editor open, ask your assistant:

• "List the actors in the current level." A read: nothing changes.
• "Spawn a point light 300 units above the origin." A write: a new actor appears.
• "Take a screenshot of the viewport." The image comes back, so a model that accepts images can check its own work.

Behind the scenes each step is up to three calls to the unreal tool: search finds a capability from a few plain words, describe returns its exact parameters, and execute runs it and returns a receipt of what changed:

{ "operation": "search", "query": "spawn actor" }
{ "operation": "describe", "tool": "control_actor", "action": "spawn" }
{
  "operation": "execute",
  "tool": "control_actor",
  "action": "spawn",
  "params": { "classPath": "/Script/Engine.PointLight", "actorName": "KeyLight", "location": [0, 0, 300] }
}

Deletes, and some other writes, only run with a consent grant that describe hands out, so they take one extra step. That is deliberate; see the security model.