Godot Agent Arena

godot-agent-arena

A Docker container that runs the Godot Engine with noVNC and Apache Guacamole, making it accessible directly from a web browser — built specifically for testing browser-based AI agents.

What is this?

This project spins up a Docker container with Godot 4.6.2 running inside a virtual display (Xvfb), exposed via two different browser access methods:

• noVNC (lightweight, websocket-based) — connect directly on port 9488
• Apache Guacamole (full remote desktop gateway) — connect on port 8080

You open a browser, connect to the container, and get a fully interactive Godot editor — no local installation needed.

Why two access methods?

Testing revealed that **the access method matters for AI agent performance**. Guacamole renders the remote session differently than noVNC, and some agents handle one much better than the other. In particular, Copilot performs significantly better when connected through Guacamole compared to noVNC.

Purpose

This is a testbed for evaluating how well **browser-based AI agents** can operate as general-purpose GUI automation tools. The agents being tested:

• Copilot (Edge)
• Comet (Perplexity)
• Claude in Chrome (Anthropic)

The workflow: start the container, open Godot in the browser, and instruct a browser agent to click through the editor and build a project from scratch.

Demo Videos

Copilot (Edge)

Guacamole

[Edge Copilot Guacamole Fail.webm](https://github.com/user-attachments/assets/f949dbdc-1b41-4f3a-92fa-576b1edeeb52)

NoVnc

[Edge Copilot NoVNC Fail.webm](https://github.com/user-attachments/assets/44a28a63-3d1f-4236-b4bd-a8ebd9f61823)

Claude in Chrome (Anthropic)

Guacamole

[Chrome Sonnet Guacamole Fail.webm](https://github.com/user-attachments/assets/623bf8ca-c37a-4ec7-b68d-3c33a79ecedf)

NoVnc

[Chrome Sonnet NoVnc Fail.webm](https://github.com/user-attachments/assets/e4450d92-cd44-4687-ae92-6c17302be0ca)

Comet (Perplexity)

Sonnet

[Comet S