Godot MCP: Game Development With an AI Agent

Godot MCP is a name for community-built Model Context Protocol servers that let an AI agent such as Claude Code, Cursor or Codex work with the Godot game engine: launch the editor, run your project, read errors from the debug output, create scenes and add nodes. As of September 30, 2026, the Godot project itself does not ship an MCP server, and none of the servers below is made or endorsed by the Godot Foundation. That is not a reason to avoid them, but it does mean you choose, vet and update them yourself, the same way you would any third-party add-on.

The main Godot MCP servers

Several projects use the name, and they work in two different ways. Some drive Godot from outside, through its command line. Others install a plugin inside the editor and talk to the live editor session. The two best-known open source options show the difference.

• Godot MCP by Coding-Solo. A Node.js server, MIT-licensed, installed with npx @coding-solo/godot-mcp. Its README↗ (opens in a new tab) lists tools to launch the editor, run a project in debug mode, capture debug output, stop a running project, list projects in a folder, get project info, create scenes, add nodes, load sprites, export a MeshLibrary and manage UIDs on Godot 4.4 and later. It needs Node.js 18 or newer and a Godot install it can find.
• Godot AI by hi-godot. An editor add-on plus a Python server, also MIT-licensed. Its README↗ (opens in a new tab) says it connects to a live editor with 46 tools covering scenes, nodes, scripts, signals, UI, materials, animation, particles, cameras and environments. Version 4 needs Godot 4.7 or later and uv, which provides uvx.
• Commercial and newer listings. Godot’s Asset Library also carries MCP add-ons such as Godot MCP Pro, which its listing↗ (opens in a new tab) marks as proprietary, for Godot 4.4 and later, with the server itself sold separately from the editor plugin.

Pick by how you work. If you mostly want an agent to run the game, read the errors and fix GDScript, the command-line server is small and easy to reason about. If you want the agent to build scenes in the editor you have open, an editor plugin gives it far more reach, and far more ways to change things you did not ask it to change.

How a command-line Godot MCP server works

Godot can do a lot without its window. The command line reference↗ (opens in a new tab) documents --path to point at a folder containing project.godot, --editor to open the editor instead of running a scene, --headless for running without a display, and --script to run a GDScript file. Coding-Solo’s server builds on that: simple operations call the Godot binary directly, and scene edits go through one bundled script, godot_operations.gd, which takes an operation name and parameters as JSON. Nothing runs inside your open editor, so a scene the agent edits is saved to disk and your editor picks it up when it reloads.

Set up Godot MCP with Claude Code or Cursor

For the Coding-Solo server, Claude Code needs one command. If Godot is not on your PATH, or you keep several versions, set GODOT_PATH to the executable you want the agent to use. Pinning a version number keeps an unreviewed update from arriving the next time npx runs.

claude mcp add godot -e GODOT_PATH=/Applications/Godot.app/Contents/MacOS/Godot -- npx -y @coding-solo/godot-mcp@0.1.1

Cursor and other clients take the same server as JSON, for example in .cursor/mcp.json in your project. Set DEBUG to "true" only while you are troubleshooting, because it makes the server log in detail.

{
  "mcpServers": {
    "godot": {
      "command": "npx",
      "args": ["-y", "@coding-solo/godot-mcp@0.1.1"],
      "env": {
        "GODOT_PATH": "C:\\Godot\\Godot_v4.7.2-stable_win64.exe"
      }
    }
  }
}

If the agent reports that Godot cannot be found, GODOT_PATH is almost always the fix. If it says the project path is invalid, check that the folder it was given contains project.godot. General connection problems are covered in Claude Code MCP not working.

Set up Godot AI in the editor

1. Download a published release, not a source snapshot, and put the add-on at your-project/addons/godot_ai/ with plugin.cfg inside that folder.
2. In Godot, open Project, then Project Settings, then Plugins, and enable Godot AI. The plugin starts the MCP server and adds a Godot AI dock.
3. In the dock, press Configure next to your client, or Configure all. The dock writes the client entry for you, with the matching version and ports. Use its command rather than a hand-written URL, because the project says a bare local HTTP entry cannot authenticate.
4. Restart the client if it does not see the new server, then start with a read: “Show me the current scene hierarchy.”

Two details from the project’s own docs are worth knowing before you commit to it. GDScript is fully supported, but C# is text only: the tools write .cs files, and you build and read compiler errors in Godot or with dotnet build. And it sends usage telemetry, which it says excludes code, scene contents and file names; setting GODOT_AI_DISABLE_TELEMETRY=true turns it off.

Cautions before you connect one

• Community means unreviewed by Godot. The Asset Library docs↗ (opens in a new tab) describe it as a repository of user-submitted add-ons, so being listed there is not an endorsement or a security review. Those docs also say every asset is free and open source, while the library now carries proprietary MCP listings, so read each listing’s license yourself.
• It runs as you. A server that can launch Godot and run scripts can do anything your user account can do on that machine. Read the tool list before you approve it, and keep approval prompts on for tools that write or run code rather than adding them to an auto-approve list.
• Pin and update on purpose. Use a release tag or a fixed npm version, read the changelog, and upgrade when you choose to.
• Match the Godot version. Godot 4.7 is the current stable series according to the Godot 4.7 release page↗ (opens in a new tab), and each server states its own minimum. A server built for one 4.x release can fail quietly on another.
• Commit before every session. An agent that edits scenes can touch .tscn, .tres and script files at once, and a diff is the only reliable record of what changed.

The wider checklist for any third-party server is in MCP security best practices, and the approval settings that decide what runs without asking are covered in Claude Code auto-approve.

A Godot dev loop with an agent

• Start with a run. “Run the project, wait five seconds, then read the debug output and summarize every error” proves the connection and gives the agent something real to fix.
• Fix one error at a time. Ask for the GDScript change, then ask it to run the project again and confirm the error is gone before moving on.
• Small scene edits. “Add a CollisionShape2D to the Player scene and save it.” Open the scene in the editor and look before asking for more.
• Assets from elsewhere. Models made in Blender come in as glTF; the modeling side is in the Blender MCP guide. If you also work in Unity, Unity MCP covers its official route, which is different from Godot’s community-only picture.

Keeping playtest bugs on a board

Errors from a run and notes from playtesters turn into work that outlives one chat. With fenbs connected as a second MCP server at https://fenbs.ai/api/mcp, you can say: “Run the project, read the debug output, and for each distinct error file a bug in fenbs with the script path and line in the note. Search the board first and comment on an existing bug instead of filing a duplicate.” Bugs land in To Do with a priority from 1 to 10, where 1 is the most urgent. When the agent picks one up, it moves it to In Progress, writes its plan into the task, and sets the test status and test notes when it is done, and the history records which assistant changed what.

fenbs is a plain task board, not a game production suite. It has no sprints, no due dates and no settable assignee, so a milestone or a tester’s name goes in the note. Rules that must hold for every session, such as “never edit files under addons/”, belong on the Decisions and rules page, which every connected assistant reads first. How that rhythm works day to day is in the Claude Code task tracking workflow.