godot-mcp-runtime

Playwright MCP for Godot, screenshots, SceneTree manipulation, and arbitrary GDScript execution at runtime through a local UDP bridge.

Godot MCP Server

npm version npm downloads License: MIT Node.js

An MCP server that gives AI assistants direct access to a running Godot 4.x game. Not just file editing, not just scene manipulation. Actual runtime control: input simulation, screenshots, UI discovery, and live GDScript execution while the game is running.

Most Godot MCP servers operate headlessly. They can create scenes, add nodes, attach scripts. That covers a lot of ground, but stops at the editor boundary. This one doesn't. When you run a project through this server, it injects a lightweight UDP bridge as an autoload, and suddenly the AI can interact with your game the same way a player would: press keys, click buttons, read what's on screen, and run arbitrary code against the live scene tree.

The distinction matters: the AI doesn't just write your game, it can check its work.

Think of it as Playwright MCP, but for Godot. Playwright lets agents verify that a web app actually works by driving a real browser. This does the same thing for games: run the project, take a screenshot, simulate input, read what's on screen, execute a script against the live scene tree. The agent closes the loop on its own changes rather than handing off to you to verify.

This is not a playtesting replacement. It doesn't catch the subtle feel issues that only a human notices, and it won't tell you if your game is fun. What it does is let an agent confirm that a scene loads, a button responds, a value updated, a script ran without errors. That's a fundamentally different development workflow, and it's what this server is built for.

This server is built around a small set of composable, wide-reaching tools rather than a long list of narrow ones. Each tool is designed to teach agents how to use it well. Response messages include next steps, timing constraints, and cleanup reminders so agents stay on track without extra prompting.

What It Does

Headless editing. Create scenes, add nodes, set properties, attach scripts, connect signals, manage UIDs, validate GDScript. All the standard operations, no editor window required.

Runtime bridge. When run_project is called, the server injects McpBridge as an autoload. This opens a UDP channel on port 9900 (localhost only) and enables:

  • Screenshots: Capture the viewport at any point during gameplay
  • Input simulation: Batched sequences of key presses, mouse clicks, mouse motion, UI element clicks by name or path, Godot action events, and timed waits
  • UI discovery: Walk the live scene tree and collect every visible Control node with its position, type, text content, and disabled state
  • Live script execution: Compile and run arbitrary GDScript with full SceneTree access while the game is running

The bridge cleans itself up automatically when stop_project is called. No leftover autoloads, no modified project files.

Quick Start

Prerequisites

Install via npm

npm install -g godot-mcp-runtime

Or clone from source

git clone https://github.com/Erodenn/godot-mcp-runtime.git
cd godot-mcp-runtime
npm install
npm run build

Configure Your MCP Client

Add the following to your MCP client config. Works with Claude Code, Claude Desktop, Cursor, or any MCP-compatible client.

If installed via npm:

{
  "mcpServers": {
    "godot": {
      "command": "godot-mcp-runtime",
      "env": {
        "GODOT_PATH": "<path-to-godot-executable>"
      }
    }
  }
}

If cloned from source:

{
  "mcpServers": {
    "godot": {
      "command": "node",
      "args": ["<path-to>/godot-mcp-runtime/dist/index.js"],
      "env": {
        "GODOT_PATH": "<path-to-godot-executable>"
      }
    }
  }
}

If Godot is on your PATH, you can omit GODOT_PATH entirely. The server will auto-detect it. Set "DEBUG": "true" in env for verbose logging.

Verify

Ask your AI assistant to call get_project_info. If it returns a Godot version string (e.g., 4.4.stable), you're connected and working.

Tools

This server intentionally keeps the tool count small. Each tool covers a broad surface area through operations and parameters, so agents spend less time finding the right tool and more time doing useful work. Tool descriptions and responses are written to guide agent behavior: what to call next, when to wait, and how to recover from errors.

Project

ToolDescription
launch_editorOpen the Godot editor GUI for a project
run_projectRun a project in debug mode and inject the MCP bridge
stop_projectStop the running project and remove the bridge
get_debug_outputRead stdout/stderr from the running project
list_projectsFind Godot projects in a directory
get_project_infoGet project metadata and Godot version

Runtime (requires run_project first)

After calling run_project, wait 2-3 seconds for the bridge to initialize before using these tools.

ToolDescription
take_screenshotCapture a PNG of the running viewport
simulate_inputSend batched input: key, mouse_button, mouse_motion, click_element, action, wait
get_ui_elementsGet all visible Control nodes with positions, types, and text
run_scriptExecute arbitrary GDScript at runtime with full SceneTree access

Scene: manage_scene

All mutation operations save automatically. Use save only for save-as (newPath) or to re-canonicalize a .tscn file.

OperationDescription
createCreate a new scene file
add_nodeAdd a node to an existing scene
load_spriteSet a texture on a Sprite2D, Sprite3D, or TextureRect
saveRe-pack and save the scene, or save-as with newPath
export_mesh_libraryExport scenes as a MeshLibrary for GridMap
batchExecute an ordered sequence of add_node, load_sprite, and save ops across one or more scenes in a single Godot process

Node: manage_node

All mutation operations save automatically.

OperationDescription
get_treeGet the full scene tree hierarchy
listList direct child nodes of a node
get_propertiesRead properties from one node, or pass a nodes array to read from multiple nodes in one process
update_propertySet a property on a node, or pass an updates array to set multiple properties in one process
attach_scriptAttach a GDScript to a node
duplicateDuplicate a node within the scene
deleteRemove a node from the scene
get_signalsList all signals on a node with their connections
connect_signalConnect a signal to a method on another node
disconnect_signalDisconnect a signal connection

Project Settings: manage_project

Edits project.godot directly, no Godot process required. Safe to use even when autoloads are broken or headless operations are failing.

OperationDescription
list_autoloadsList all registered autoloads with paths and singleton status
add_autoloadRegister a new autoload
remove_autoloadUnregister an autoload by name
update_autoloadModify an existing autoload's path or singleton flag
get_project_settingsRead settings from project.godot by section and key
get_filesystem_treeGet the project file tree with types and extensions
search_in_filesSearch for a string across project source files
get_scene_dependenciesList all resources a scene depends on

Validate: validate

Validate before attaching or running. Catches syntax errors and missing resource references before they cause headless crashes or runtime failures.

InputDescription
scriptPathValidate an existing .gd file in the project
sourceValidate inline GDScript written to a temp file
scenePathValidate a .tscn file and check that all ext_resource references resolve
targetsArray of the above — validate multiple scripts or scenes in a single Godot process

Returns { valid, errors: [{ line?, message }] } per target. Fix reported errors and re-validate before calling attach_script or run_script.

UIDs: manage_uids (Godot 4.4+)

OperationDescription
getGet a resource's UID
updateResave all resources to update UID references

Architecture

src/
├── index.ts                # MCP server entry point, routes tool calls
├── tools/
│   ├── project-tools.ts    # Project and runtime tool handlers
│   ├── scene-tools.ts      # Scene operations
│   ├── node-tools.ts       # Node operations
│   └── validate-tools.ts   # GDScript and scene validation
├── scripts/
│   ├── godot_operations.gd # Headless GDScript operations
│   └── mcp_bridge.gd       # UDP autoload for runtime communication
└── utils/
    └── godot-runner.ts     # Process spawning, output parsing

Headless operations spawn Godot with --headless --script godot_operations.gd, perform the operation, and return JSON. Runtime operations communicate over UDP with the injected McpBridge autoload.

How the Bridge Works

When run_project is called:

  1. mcp_bridge.gd is copied into the project directory
  2. It's registered as an autoload in project.godot
  3. The project launches with the bridge listening on 127.0.0.1:9900
  4. Runtime tools send JSON commands to the bridge and await responses
  5. When stop_project is called, the autoload entry and bridge script are removed

Files generated during runtime (screenshots, executed scripts) are stored in .mcp/ inside the project directory. This directory is automatically added to .gitignore and has a .gdignore so Godot won't import it.

Broken Autoloads

If any registered autoload fails to initialize (syntax error, missing resource, display dependency), Godot's headless process will crash before any operation runs. Use manage_project to inspect and remove the failing autoload. It edits project.godot directly, with no Godot process involved.

Acknowledgments

Built on the foundation laid by Coding-Solo/godot-mcp for headless Godot operations.

Developed with Claude Code.

License

MIT

Related Servers