𝗧𝗲𝘀𝘁𝗶𝗻𝗴 𝗚𝗼𝗱𝗼𝘁 𝗠𝗖𝗣𝘀 𝘄𝗶𝘁𝗵 𝗮 𝗠𝗶𝗴𝗿𝗮𝘁𝗶𝗼𝗻 𝗧𝗲𝘀𝘁
I tested three Godot MCPs. I gave Codex the same task: migrate a Godot 3.6 project to Godot 4.6.
The project was messy. It had 208 errors and 73 warnings. This was a real test, not a small demo.
I tested Fennara (my own MCP), Godot AI, and Godot MCP Native.
Here is what I found:
Fennara Results Fennara focused on deep feedback. It did more than read editor output. It surfaced:
- Script errors
- Scene errors
- Shader errors
- Old API errors
Codex used the get_class_info tool often. This prevented the AI from guessing new APIs. When Codex edited a file, Fennara returned new errors immediately. The AI did not work blindly. It received instant feedback from Godot.
Godot AI Results This tool struggled with feedback.
- It reported the project was running even when the scene was broken.
- The editor logs contained stale errors from old file versions.
- Codex tried to fix errors that no longer existed.
- The AI thought it finished, but 30 errors remained after a reload.
Godot MCP Native Results This tool had similar issues to Godot AI.
- Having too many tools made it hard for the agent to choose the right one.
- Codex relied on the terminal instead of the MCP tools.
- Like the others, it left errors behind after claiming it was done.
My Conclusion An AI agent needs more than file access. It needs a reliable feedback loop.
The biggest problems in Godot AI agents are:
- Stale editor output
- Unclear runtime errors
- Editor cache issues
- Fixing errors that are already gone
- Too many tools causing confusion
The best MCP is not the one with the most commands. It is the one that helps the AI understand what Godot is actually saying. Reliable feedback stops the AI from hallucinating.
Optional learning community: https://t.me/GyaanSetuAi