Spielerando

Godot AI MCP: Let Your AI Run the Game and Read Errors

Godot AI MCP means connecting your AI tool to Godot through a small add-on, so the AI can run your game, read the errors and change scenes by itself. That removes the slowest part of vibe coding, where you copy error messages back and forth by hand.

If MCP is new to you, read what MCP is first. It also lists six questions to ask before you connect any server.

Which server should I look at?

Here are two, as described on their GitHub pages on 2 November 2026. Others exist, and this area moves fast.

godot-mcp GDAI MCP
Runs the game, reads errors Yes: runs it and fetches debug output Reads debugger output, logs and script errors
Sees the screen Not listed Takes screenshots of the editor and running game
Changes scenes Creates scenes, adds parts, loads pictures Yes, by its description
You need Godot, Node.js 18 or newer, npm Godot 4.1 or newer
Licence MIT, source code public "All rights reserved", licence on its site
Where from Its GitHub page, through npm Its own site, gdaimcp.com

The choice comes down to eyes and openness. If your bugs are mostly crashes, godot-mcp is enough, and you can read all of its code. If your bugs are often things that look wrong, screenshots matter, so look at GDAI MCP and read its licence and price on its own site first.

To add godot-mcp to Claude Code, run this:

claude mcp add godot -- npx @coding-solo/godot-mcp

Save a working copy with Git before you connect anything. A server that can change scenes can also break them.

What changes in practice

Sandcastle Siege is an invented game where the tide rises and you defend a sandcastle. It crashes when the water reaches the outer wall.

Without MCP, you start the game, wait for the tide, read the error, copy it into a chat, paste the fix, start again. Each round takes minutes, and you are only a messenger.

With MCP, the AI runs the game, reads the error, changes the scene and runs it again. Your job moves from messenger to judge. You also get a new risk, because an AI that can try things can also try things twenty times and declare victory. So ask it to explain the crash before it touches anything. It can read the real scene from the project, so it can draw the chain of events from facts, not memory. That drawing is in Mermaid, a text format that mermaid.live displays.

MCP can also carry knowledge, not only actions. Spielerando is one such connection: your AI gets blueprints, ready-made code and a wiki. You need a free account with Google or e-mail, and it does not look at your game or code. See the setup page and the plans.

Paste this to your AI

You are connected to my Godot project through the Godot MCP server. The game is Sandcastle Siege, and it crashes when the tide reaches the outer wall.

Run the project, reach that moment if you can, and read the debug output. Quote the exact error line before you change anything.

Then draw the chain of events as a Mermaid diagram, from the project as it is: the tide rises, what it touches, what each step calls. Mark the step where the error happens. Check that the diagram fits both the error and the scene you read. If they disagree, say so and keep looking.

Make the smallest change that follows from the diagram, run the project again and quote the new output. If you could not reach the crash, say so and tell me how to reach it by hand. Do not claim a fix that rests only on reading the code.

Or skip this and connect the MCP: your AI then finds all of this by itself.Connect the MCP

How to check it worked

  • Find the quoted error in Godot's own output panel. It should be there, word for word.
  • Let the tide rise three times yourself. One clean run says little in a game with timing.
  • Look at what changed. Ask for the list of files. Anything unrelated to the crash needs a reason.
  • Turn the server off and ask it to run the game. It should tell you that it cannot.

Sources

Last updated: 2026-11-02