Model Context Protocol lets an AI assistant work against your actual Godot project
instead of guessing from pasted code. Here is what it is, how to set it up, and where
the sharp edges are.
Godot ships no official MCP server. Everything below is community tooling, and the
setup differs depending on whether the server runs outside the editor or inside it.
What MCP actually is
MCP is an open standard for connecting AI applications to external systems. It defines
three roles: a host (the AI app, such as Claude Code or Cursor), a client the host creates for each connection, and a server that exposes capabilities. Servers can offer tools (functions the model can call), resources (data it can read), and prompts (reusable templates).
Local servers talk over stdio - a plain process on your machine, no
network. Remote servers use streamable HTTP. Every Godot MCP server you
are likely to install runs locally over stdio, which is why setup is just a command and
a config file rather than an account and an API key.
Why bother, versus pasting code into a chat
Pasting code gives the model a snapshot with no way to check itself. A connected server
closes that loop: the assistant can run the project, read the actual error, change a
scene, and run it again. The practical difference shows up in debugging, where the
model stops guessing what your stack trace probably said.
It does not make the model a better Godot programmer. It makes it an accountable one.
For the language and API knowledge itself, a good system prompt still matters more -
see the Godot Developer Agent file.
Two architectures, two different tools
External process server
godot-mcp, and most Node/TypeScript servers
Runs as a separate Node process. Drives Godot through its command line, so it can launch the editor, run the project headlessly, and read the console.
No addon inside your project
Works without the editor open
Easy to add or remove per client
Cannot see live editor state
Scene edits happen through files, not the open editor
In-editor plugin server
godot-ai, Godot-MCP-Native, gdai-mcp-plugin-godot
Ships as a GDScript addon inside your project and talks to the client from within a running editor session.
Sees the live scene tree and selection
Can use editor-only APIs
Immediate feedback in the editor
The editor must be open
An addon lives in your project
Tied to a Godot version's plugin API
They are not competitors. An external server is better at "run this and tell me what
broke"; an in-editor plugin is better at "look at what I have selected right now".
Set it up
Prerequisites: Godot installed, Node.js 18 or newer, and an MCP-capable
client. The examples use godot-mcp, the most widely used external-process
server.
One command. Claude Code stores the server in its own MCP config for you.
claude mcp add godot -- npx @coding-solo/godot-mcp
Restart the client afterwards. Most hosts read their MCP config only at startup, which
is the single most common reason a freshly added server shows no tools.
What you can actually do with it
These are the tools godot-mcp exposes, grouped by how much damage they can
do. The grouping matters more than the names: it is the basis for deciding what to
auto-approve.
Inspect
Read-only. Safe to auto-approve.
get_godot_version
Report the installed Godot version.
list_projects
Find Godot projects in a directory.
get_project_info
Read project structure and metadata.
get_uid
Read a file's UID (Godot 4.4+).
Run and debug
Launches processes on your machine.
launch_editor
Open the Godot editor for a project.
run_project
Run the project in debug mode.
get_debug_output
Read console output and errors.
stop_project
Stop the running project.
Edit
Writes to disk. Review these before approving.
create_scene
Create a scene with a given root node type.
add_node
Add a node to a scene with properties.
load_sprite
Load a texture into a Sprite2D.
save_scene
Save a scene, optionally as a variant.
export_mesh_library
Export a 3D scene as a MeshLibrary.
update_project_uids
Resave resources to refresh UID references.
A realistic first session
Once connected, the useful prompts are the ones that ask the assistant to verify, not just write:
"Check my Godot version and list the projects under my dev folder."
"Run the project and show me the debug output."
"The player falls through the floor. Run it, read the errors, and tell me what you find."
"Create a Player scene with a CharacterBody2D root, a Sprite2D, and a CollisionShape2D."
The fourth one writes to disk. The first three do not. That distinction is the whole
safety model.
Safety and limits
Commit first. An agent editing scenes is a script with write access. Version control is your undo.
Auto-approve reads only. Version, project listing, project info, and debug output are safe. Scene edits are not.
A real project is required. Tools operate on a directory containing project.godot.
UID tools need Godot 4.4+.get_uid and update_project_uids do not apply to older projects.
It will not fix bad Godot knowledge. A model that writes Godot 3 APIs will still write them, just faster. See using AI with Godot for how to handle that.
Troubleshooting
The client lists no Godot tools
Restart the client after editing its config. Most MCP hosts only read server config at startup.
"Godot executable not found"
Set GODOT_PATH in the server's env block to the absolute path of your Godot binary. Automatic detection misses Steam and custom installs.
Tools fail with a project error
Point the tool at a directory that contains a project.godot file. The server operates on real projects, not arbitrary folders.
get_uid or update_project_uids errors
UIDs exist from Godot 4.4 onward. On older projects these two tools do not apply.
npx re-downloads the server every run
Install the package globally, or clone and build from source and point command at the built entry file.
Godot MCP: common questions
What is MCP in Godot?
MCP (Model Context Protocol) is an open standard that lets an AI application talk to external tools. A Godot MCP server exposes Godot operations - run the project, read the debug console, create a scene, add a node - as tools your AI client can call, so the assistant works against your real project instead of guessing from pasted snippets.
How do I set up a Godot MCP server?
Install Node.js 18 or newer and Godot, then register the server with your AI client. In Claude Code the command is: claude mcp add godot -- npx @coding-solo/godot-mcp. In Cursor or Cline you add an mcpServers entry to the client's JSON config with the same command and args, then restart the client.
Does Godot have official MCP support?
No. Godot does not ship an MCP server in the engine. Every Godot MCP server today is a community project, either an external process that drives the Godot command line or a GDScript addon that runs inside the editor.
Is it safe to let an AI edit my Godot project through MCP?
Treat it like giving a script write access to your project folder. Keep the project in version control, commit before an agent session, and only auto-approve read-only tools such as get_godot_version, list_projects, get_project_info, and get_debug_output. Review scene-editing and file-writing calls before approving them.
Which Godot MCP server should I use?
If you want the AI to run your project and read errors without keeping the editor open, use an external-process server such as godot-mcp. If you want the AI to see the live scene tree and selection inside an open editor, use an in-editor GDScript plugin server. The two approaches solve different problems and can be used together.