AI Playtester

Godot Plugin UI for user objectives and api connections
AI actions
web server
Ending short summary
full detailed summary of playtest
AI playtesting my game, what???

Inspiration

We as indie game developers want to develop more features for a game but sometimes it is too tedious to continually play the game to test if the new feature will work. Since there are newer AI agents that can code and work autonomously, we had the idea to build an AI agent that can assist game developers to play test games their games for them.

What it does

The AI agent will have access to your Godot game engine in the form of a plugin. The developer can input objectives that the agent can fulfill. For example, I made a sword and I wanted to see if my player could be able to pick up that sword so I give the objective of "pick up sword" to the agent. The agent will play the game to pick up the sword.

How we built it

We built it using .NET 10, Godot 4.4, Claude code and Gemini 3.0 for initial code base for the plugin and web server. We configured and debugged some issues that it had with our demo computer. We used C# to build a web server that could act as the bridge between Claude Console and Godot through a websocket. The core engine of the server allows the agent to observer, think, act and it will loop through that until it completes its objective.

Challenges we ran into

Challenges included being able to get the plugin to work on Godot. Debugging in a language we were not familiar with, godot script and c#. The Web server was not responding and connecting to the Claude Console. The AI agent did not complete objectives or achieve desired behaviors. The UI was not user friendly nor showing the relevant information for the developer.

Accomplishments that we're proud of

That we were able to get a concept from our mind into reality. We are proud that the agent could fulfill objectives and have a very close desirable outcomes. That the plugin and ai agent worked.

What we learned

(Ethan) I learned how to effectively use AI, through effective prompting and how to work around weaknesses such as specific and detail prompting. The AI would make incomplete progress so I had to make sure that the prompts were effective. (Tony) I learned how tedious training a model could be. While we are using a model that holds billions of parameter sizes, this product could be used more of a way to fine tune a local AI agent to play the game for you through lots of trial and errors. I learned how to make a plugin for Godot.