An AI Agent Built a Complete 3D Tower Defense in Godot — Here's How

An AI coding agent just built a playable 3D tower defense from an empty Godot project. No human coding.

I built GodotIQ - an MCP server with 36 tools that gives AI agents spatial intelligence inside Godot 4. To stress-test it, I pointed Claude Code at an empty project with a game design doc and the Kenney Tower Defense Kit.

The agent built everything autonomously:

The game — "Crystal Keep":

• 12×10 terrain map with connected S-shaped path
• 4 tower types: Gatling (fast DPS), Cannon (splash damage), Frost (slows enemies), Support (buffs nearby towers)
• 3 upgrade levels per tower, sell with 60% refund
• 4 enemy types including bosses every 5 waves
• 20 waves of increasing difficulty
• Full placement system — click tower card, click on the map, tower appears
• Tower selection panel with upgrade/sell buttons
• Camera controls (WASD pan, scroll zoom, Q/E rotate)
• Pause (P/Space) and speed control (1x/2x/3x)
• Main menu, tutorial, game over screen with stats and score
• Visual effects: screen shake, damage numbers, gold popups, crystal animation

What makes this different

Most AI tools for Godot either wrap the editor API or generate code from prompts. The agent writes scripts, but it has no idea what the 3D scene actually looks like.

GodotIQ adds an intelligence layer. The agent can:

• See the 3D space — query a spatial map with positions, distances, bounds of every object
• Place objects intelligently — validate positions against grid, walls, navmesh, other objects
• Interact with the game — click buttons, click positions in the 3D viewport, simulate player input
• Verify its own work — take screenshots from editor and running game, inspect node states at runtime
• Understand the code — trace signal flows, map dependencies, check impact of changes before making them

The agent doesn't just write code. It develops the game — builds scenes, tests them, sees the result, fixes what's wrong, and iterates.

The workflow

The agent followed this loop hundreds of times:

1. Build — create scene nodes, write scripts
2. Verify — take screenshot, inspect state
3. Fix — if something looks wrong, patch it
4. Test — run the game, simulate player actions, check results

For tower placement, the agent used click_at_world to simulate clicking on 3D positions — testing the placement system exactly as a player would. For UI, it used click_at to press buttons and verify they worked. No shortcuts, no bypassing the game's input system.

Numbers

• 36 tools across 9 categories (spatial, code, bridge, flow, assets, memory, UI, animation, navigation)
• 22 tools free (spatial intelligence, code analysis, flow tracing)
• 14 PRO tools ($19 one-time) for runtime features (screenshots, game control, input simulation)
• 1460 automated tests
• Works with Claude Code, Cursor, Windsurf, GitHub Copilot, Codex, and any MCP client

Try it

bash
pip install godotiq

Website: godotiq.com
GitHub: github.com/salvo10f/godotiq

The game assets are from the Kenney Tower Defense Kit (CC0, free).

GodotIQ is open-core. The intelligence layer (spatial understanding, code analysis, flow tracing) is free. Runtime tools (screenshots, game control, input sim) require a $19 one-time Pro license.