Claude spawning scene objects and controlling their transformations and materials, generating blueprints, functions, variables, adding components, running python scripts etc.

A project called become human, where NPCs are OpenAI agentic instances. Built using this plugin.

[!WARNING]
This plugin is still under rapid development.

1. Do not use it in production environments. ⚠️
2. Do not use it without version control. ⚠️

A stable version will be released soon. 🚀🔥

Every month, hundreds of new AI models are released by various organizations, making it hard to keep up with the latest advancements.

The Unreal Engine Generative AI Support Plugin allows you to focus on game development without worrying about the LLM/GenAI integration layer.

Currently integrating Model Control Protocol (MCP) with Unreal Engine 5.5.

This project aims to build a long-term support (LTS) plugin for various cutting-edge LLM/GenAI models and foster a community around it. It currently includes OpenAI's GPT-4o, Deepseek R1, Claude Sonnet 3.7 and GPT-4o-mini for Unreal Engine 5.1 or higher, with plans to add , real-time APIs, Gemini, MCP, and Grok 3 APIs soon. The plugin will focus exclusively on APIs useful for game development, evals and interactive experiences. All suggestions and contributions are welcome. The plugin can also be used for setting up new evals and ways to compare models in game battlefields.

• OpenAI API Support:
  • OpenAI Chat API ✅ (models-ref)
    • gpt-4o, gpt-4o-mini Model ✅
    • gpt-4.5-preview Model 🛠️
    • o1-mini, o1, o1-pro Model 🚧
    • o3-mini Model 🛠️
  • OpenAI DALL-E API ❌ (Until new generation models are released)
  • OpenAI Vision API 🚧
  • OpenAI Realtime API 🛠️
    • gpt-4o-realtime-preview gpt-4o-mini-realtime-preview Model 🛠️
  • OpenAI Structured Outputs ✅
  • OpenAI Whisper API 🚧
• Anthropic Claude API Support:
  • Claude Chat API ✅
    • claude-3-7-sonnet-latest Model ✅
    • claude-3-5-sonnet Model ✅
    • claude-3-5-haiku-latest Model ✅
    • claude-3-opus-latest Model ✅
  • Claude Vision API 🚧
• XAI (Grok 3) API Support:
  • XAI Chat Completions API 🚧
    • grok-beta Model 🚧
    • grok-beta Streaming API 🚧
  • XAI Image API 🚧
• Google Gemini API Support:
  • Gemini Chat API 🚧🤝
    • gemini-2.0-flash-lite, gemini-2.0-flash gemini-1.5-flash Model 🚧🤝
  • Gemini Imagen API: 🚧
    • imagen-3.0-generate-002 Model 🚧
• Meta AI API Support:
  • Llama Chat API ❌ (Until new generation models are released)
    • llama3.3-70b Model ❌
    • llama3.1-8b Model ❌
  • Local Llama API 🚧🤝
• Deepseek API Support:
  • Deepseek Chat API ✅
    • deepseek-chat (DeepSeek-V3) Model ✅
  • Deepseek Reasoning API, R1 ✅
    • deepseek-reasoning-r1 Model ✅
    • deepseek-reasoning-r1 CoT Streaming ❌
  • Independently Hosted Deepseek Models 🚧
• Baidu API Support:
  • Baidu Chat API 🚧
    • baidu-chat Model 🚧
• 3D generative model APIs:
  • TripoSR by StabilityAI 🚧
• Plugin Documentation 🛠️🤝
• Plugin Example Project 🛠️ here
• Version Control Support
  • Perforce Support 🚧
  • Git Submodule Support ✅
• LTS Branching 🚧
  • Stable Branch with Bug Fixes 🚧
  • Dedicated Contributor for LTS 🚧
• Lightweight Plugin (In Builds)
  • No External Dependencies ✅
  • Build Flags to enable/disable APIs 🚧
  • Submodules per API Organization 🚧
  • Exclude MCP from build 🚧
• Testing
  • Automated Testing 🚧
  • Different Platforms 🚧🤝
  • Different Engine Versions 🚧🤝

• Clients Support ✅
  • Claude Desktop App Support ✅
  • Cursor IDE Support ✅
  • OpenAI Operator API Support 🚧
• Blueprints Auto Generation 🛠️
  • Creating new blueprint of types ✅
  • Adding new functions, function/blueprint variables ✅
  • Adding nodes and connections 🛠️ (buggy)
  • Advanced Blueprints Generation 🛠️
• Level/Scene Control for LLMs 🛠️
  • Spawning Objects and Shapes ✅
  • Moving, rotating and scaling objects ✅
  • Changing materials and color ✅
  • Advanced scene features 🛠️
• Generative AI:
  • Prompt to 3D model fetch and spawn 🛠️
• Control:
  • Ability to run Python scripts ✅
  • Ability to run Console Commands ✅
• UI:
  • Widgets generation 🛠️
  • UI Blueprint generation 🛠️
• Project Files:
  • Create/Edit project files/folders ️✅
  • Delete existing project files ❌
• Others:
  • Project Cleanup 🛠️

Where,

• ✅ - Completed
• 🛠️ - In Progress
• 🚧 - Planned
• 🤝 - Need Contributors
• ❌ - Won't Support For Now

• Setting API Keys
  • For Editor
  • For Packaged Builds
• Setting up MCP
• Adding the plugin to your project
  • With Git
  • With Perforce
  • With Unreal Marketplace
• Fetching the Latest Plugin Changes
  • With Git
  • With Perforce
• Usage
  • OpenAI
    • 1. Chat
    • 2. Structured Outputs
  • DeepSeek API
    • 1. Chat and Reasoning
  • Anthropic API
    • 1. Chat
  • Model Control Protocol (MCP)
• Known Issues
• Contribution Guidelines
  • Setting up for Development
  • Project Structure
• References

[!NOTE]
There is no need to set the API key for testing the MCP features in Claude app. Anthropic key only needed for Claude API.

Set the environment variable PS_<ORGNAME> to your API key.

setx PS_<ORGNAME> "your api key"

1. Run the following command in your terminal, replacing yourkey with your API key.

   echo "export PS_<ORGNAME>='yourkey'" >> ~/.zshrc

2. Update the shell with the new variable:

   source ~/.zshrc

PS: Don't forget to restart the Editor and ALSO the connected IDE after setting the environment variable.

Where <ORGNAME> can be: PS_OPENAIAPIKEY, PS_DEEPSEEKAPIKEY, PS_ANTHROPICAPIKEY, PS_METAAPIKEY, PS_GOOGLEAPIKEY etc.

Storing API keys in packaged builds is a security risk. This is what the OpenAI API documentation says about it:

"Exposing your OpenAI API key in client-side environments like browsers or mobile apps allows malicious users to take that key and make requests on your behalf – which may lead to unexpected charges or compromise of certain account data. Requests should always be routed through your own backend server where you can keep your API key secure."

Read more about it here.

For test builds you can call the GenSecureKey::SetGenAIApiKeyRuntime either in c++ or blueprints function with your API key in the packaged build.

[!NOTE]
If your project only uses the LLM APIs and not the MCP, you can skip this section.

[!CAUTION]
Discalimer: If you are using the MCP feature of the plugin, it will directly let the Claude Desktop App control your Unreal Engine project. Make sure you are aware of the security risks and only use it in a controlled environment.

Please backup your project before using the MCP feature and use version control to track changes.

• Claude Desktop App from here.
• Cursor IDE from here.

claude_desktop_config.json file in Claude Desktop App's installation directory. (might ask claude where its located for your platform!) The file will look something like this:

{
    "mcpServers": {
      "unreal-handshake": {
        "command": "python",
        "args": ["<your_project_directoy_path>/Plugins/GenerativeAISupport/Content/Python/mcp_server.py"],
        "env": {
          "UNREAL_HOST": "localhost",
          "UNREAL_PORT": "9877" 
        }
      }
    }
}

.cursor/mcp.json file in your project directory. The file will look something like this:

{
    "mcpServers": {
      "unreal-handshake": {
        "command": "python",
        "args": ["<your_project_directoy_path>/Plugins/GenerativeAISupport/Content/Python/mcp_server.py"],
        "env": {
          "UNREAL_HOST": "localhost",
          "UNREAL_PORT": "9877" 
        }
      }
    }
}

pip install mcp[cli]

1. Add the Plugin Repository as a Submodule in your project's repository.

   git submodule add https://github.com/prajwalshettydev/UnrealGenAISupport Plugins/GenerativeAISupport

2. Regenerate Project Files: Right-click your .uproject file and select Generate Visual Studio project files.

3. Enable the Plugin in Unreal Editor: Open your project in Unreal Editor. Go to Edit > Plugins. Search for the Plugin in the list and enable it.

4. For Unreal C++ Projects, include the Plugin's module in your project's Build.cs file:

   PrivateDependencyModuleNames.AddRange(new string[] { "GenerativeAISupport" });

Still in development..

Coming soon, for free, in the Unreal Engine Marketplace.

you can pull the latest changes with:

cd Plugins/GenerativeAISupport
git pull origin main

Or update all submodules in the project:

git submodule update --recursive --remote

Still in development..

There is a example Unreal project that already implements the plugin. You can find it here.

Currently the plugin supports Chat and Structured Outputs from OpenAI API. Both for C++ and Blueprints. Tested models are gpt-4o, gpt-4o-mini, gpt-4.5, o1-mini, o1, o3-mini-high.

    void SomeDebugSubsystem::CallGPT(const FString& Prompt, 
        const TFunction<void(const FString&, const FString&, bool)>& Callback)
    {
        FGenChatSettings ChatSettings;
        ChatSettings.Model = TEXT("gpt-4o-mini");
        ChatSettings.MaxTokens = 500;
        ChatSettings.Messages.Add(FGenChatMessage{ TEXT("system"), Prompt });
    
        FOnChatCompletionResponse OnComplete = FOnChatCompletionResponse::CreateLambda(
            [Callback](const FString& Response, const FString& ErrorMessage, bool bSuccess)
        {
            Callback(Response, ErrorMessage, bSuccess);
        });
    
        UGenOAIChat::SendChatRequest(ChatSettings, OnComplete);
    }

Sending a custom schema json directly to function call

FString MySchemaJson = R"({
"type": "object",
"properties": {
    "count": {
        "type": "integer",
        "description": "The total number of users."
    },
    "users": {
        "type": "array",
        "items": {
            "type": "object",
            "properties": {
                "name": { "type": "string", "description": "The user's name." },
                "heading_to": { "type": "string", "description": "The user's destination." }
            },
            "required": ["name", "role", "age", "heading_to"]
        }
    }
},
"required": ["count", "users"]
})";

UGenAISchemaService::RequestStructuredOutput(
    TEXT("Generate a list of users and their details"),
    MySchemaJson,
    [](const FString& Response, const FString& Error, bool Success) {
       if (Success)
       {
           UE_LOG(LogTemp, Log, TEXT("Structured Output: %s"), *Response);
       }
       else
       {
           UE_LOG(LogTemp, Error, TEXT("Error: %s"), *Error);
       }
    }
);

Sending a custom schema json from a file

#include "Misc/FileHelper.h"
#include "Misc/Paths.h"
FString SchemaFilePath = FPaths::Combine(
    FPaths::ProjectDir(),
    TEXT("Source/:ProjectName/Public/AIPrompts/SomeSchema.json")
);

FString MySchemaJson;
if (FFileHelper::LoadFileToString(MySchemaJson, *SchemaFilePath))
{
    UGenAISchemaService::RequestStructuredOutput(
        TEXT("Generate a list of users and their details"),
        MySchemaJson,
        [](const FString& Response, const FString& Error, bool Success) {
           if (Success)
           {
               UE_LOG(LogTemp, Log, TEXT("Structured Output: %s"), *Response);
           }
           else
           {
               UE_LOG(LogTemp, Error, TEXT("Error: %s"), *Error);
           }
        }
    );
}

Currently the plugin supports Chat and Reasoning from DeepSeek API. Both for C++ and Blueprints. Points to note:

• System messages are currently mandatory for the reasoning model. API otherwise seems to return null
• Also, from the documentation: "Please note that if the reasoning_content field is included in the sequence of input messages, the API will return a 400 error. Read more about it here"

[!WARNING]
While using the R1 reasoning model, make sure the Unreal's HTTP timeouts are not the default values at 30 seconds. As these API calls can take longer than 30 seconds to respond. Simply setting the HttpRequest->SetTimeout(<N Seconds>); is not enough So the following lines need to be added to your project's DefaultEngine.ini file:

[HTTP]
HttpConnectionTimeout=180
HttpReceiveTimeout=180

 FGenDSeekChatSettings ReasoningSettings;
 ReasoningSettings.Model = EDeepSeekModels::Reasoner; // or EDeepSeekModels::Chat for Chat API
 ReasoningSettings.MaxTokens = 100;
 ReasoningSettings.Messages.Add(FGenChatMessage{TEXT("system"), TEXT("You are a helpful assistant.")});
 ReasoningSettings.Messages.Add(FGenChatMessage{TEXT("user"), TEXT("9.11 and 9.8, which is greater?")});
 ReasoningSettings.bStreamResponse = false;
 UGenDSeekChat::SendChatRequest(
     ReasoningSettings,
     FOnDSeekChatCompletionResponse::CreateLambda(
         [this](const FString& Response, const FString& ErrorMessage, bool bSuccess)
         {
             if (!UTHelper::IsContextStillValid(this))
             {
                 return;
             }

             // Log response details regardless of success
             UE_LOG(LogTemp, Warning, TEXT("DeepSeek Reasoning Response Received - Success: %d"), bSuccess);
             UE_LOG(LogTemp, Warning, TEXT("Response: %s"), *Response);
             if (!ErrorMessage.IsEmpty())
             {
                 UE_LOG(LogTemp, Error, TEXT("Error Message: %s"), *ErrorMessage);
             }
         })
 );

Currently the plugin supports Chat from Anthropic API. Both for C++ and Blueprints. Tested models are claude-3-7-sonnet-latest, claude-3-5-sonnet, claude-3-5-haiku-latest, claude-3-opus-latest.

    // ---- Claude Chat Test ----
    FGenClaudeChatSettings ChatSettings;
    ChatSettings.Model = EClaudeModels::Claude_3_7_Sonnet; // Use Claude 3.7 Sonnet model
    ChatSettings.MaxTokens = 4096;
    ChatSettings.Temperature = 0.7f;
    ChatSettings.Messages.Add(FGenChatMessage{TEXT("system"), TEXT("You are a helpful assistant.")});
    ChatSettings.Messages.Add(FGenChatMessage{TEXT("user"), TEXT("What is the capital of France?")});
    
    UGenClaudeChat::SendChatRequest(
        ChatSettings,
        FOnClaudeChatCompletionResponse::CreateLambda(
            [this](const FString& Response, const FString& ErrorMessage, bool bSuccess)
            {
                if (!UTHelper::IsContextStillValid(this))
                {
                    return;
                }
    
                if (bSuccess)
                {
                    UE_LOG(LogTemp, Warning, TEXT("Claude Chat Response: %s"), *Response);
                }
                else
                {
                    UE_LOG(LogTemp, Error, TEXT("Claude Chat Error: %s"), *ErrorMessage);
                }
            })
    );

This is currently work in progress. The plugin supports various clients like Claude Desktop App, Cursor etc.

Running the MCP server:

python <your_project_directoy>/Plugins/GenerativeAISupport/Content/Python/mcp_server.py

Tools -> Run Python Script -> Select the Plugins/GenerativeAISupport/Content/Python/unreal_socket_server.py file.

• Nodes fail to connect properly with MCP
• No undo redo support for MCP
• No streaming support for Deepseek reasoning model
• No complex material generation support for the create material tool
• Issues with running some llm generated valid python scripts
• When LLM compiles a blueprint no proper error handling in its response
• Issues spawning certain nodes, especially with getters and setters
• Doesn't open the right context window during scene and project files edit.
• Doesn't dock the window properly in the editor for blueprints.

1. Install unreal python package and setup the IDE's python interpreter for proper intellisense.

pip install unreal

More details will be added soon.

More details will be added soon.

• Env Var set logic from: OpenAI-Api-Unreal by KellanM
• MCP Server inspiration from: Blender-MCP by ahujasid

• OpenAI API Documentation
• Anthropic API Documentation
• XAI API Documentation
• Google Gemini API Documentation
• Meta AI API Documentation
• Deepseek API Documentation
• Model Control Protocol (MCP) Documentation
• TripoSt Documentation