Connect an AI agent to Godot or Roblox Studio with MCP

MCP (Model Context Protocol) is the open standard most AI coding tools use to control other apps. Hook it up to a game editor and your agent can read your scene, add nodes or parts, write scripts and run the game, instead of just suggesting code for you to paste.

Godot and Roblox Studio take different routes:

Roblox Studio has an official MCP server built in. You switch it on and connect your client.
Godot has no official MCP server. The most popular option is a free, MIT-licensed community plugin called Godot AI (Unofficial: it isn't made by the Godot project).

Part 1 covers Godot, Part 2 covers Roblox. Both work with Claude Code, Claude Desktop, Codex and Cursor, among others. For how Godot compares with Unity and Unreal for agent work, see Godot vs Unity vs Unreal for AI coding agents. Setup guides for the other engines: Unity and Unreal Engine 5.8.

Part 1: Godot

What you need

Godot 4.7 or newer. Godot AI v4 needs Godot 4.7+ (README). The current stable Godot is 4.7.2, released August 18, 2026 (Godot releases).
uv, a Python tool runner. The plugin uses its uvx command to start its server (README).
An MCP client: Claude Code, Claude Desktop, Codex, Cursor, VS Code, Grok Build, Gemini CLI, Zed and others.

Step-by-step

1. Install uv. On macOS or Linux (uv docs):

curl -LsSf https://astral.sh/uv/install.sh | sh

On Windows (PowerShell):

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Open a new terminal and check that uvx --version works.

2. Download the Godot AI plugin (Unofficial). Go to the Godot AI releases page. The latest release on October 18 was v4.3.0 (October 3, 2026). For a new install, download godot-ai-v4-plugin.zip. The other zip on the page, godot-ai-plugin.zip, is the upgrade package for people moving from v3.

The release page also has a manifest file (godot-ai-v4-plugin.manifest.json) that lists the zip's SHA-256 checksum, so you can check your download. On Linux or macOS, run sha256sum godot-ai-v4-plugin.zip (macOS: shasum -a 256) and compare.

3. Put it in your project. Unzip it into the root of your Godot project. You should end up with your-project/addons/godot_ai/plugin.cfg. If the plugin doesn't show up in the next step, check that the file isn't at addons/plugin.cfg instead (README).

4. Turn it on. In Godot, open Project → Project Settings → Plugins and enable Godot AI. A Godot AI dock appears. It starts the MCP server for you and should say Server connected.

5. Connect your AI client. In the dock, click Clients & Tools (or Configure an AI client), then click Configure next to your client, or Configure all. Restart the client if it doesn't notice the change (README).

Use the button rather than typing a config by hand. The plugin writes a pinned command (godot-ai attach over stdio). The README warns that a bare http://127.0.0.1:8000/mcp entry won't work with v4, because the connection is authenticated.
Claude Code also needs a one-time approval: run claude inside the project folder after configuring.
Configs are global by default. To keep one per project, set Editor Settings → Plugins → godot_ai/mcp_client_scope to project and press Configure again.

6. Give the agent its first job. The plugin's README suggests prompts like these:

Show me the current scene hierarchy.

Create a Camera3D named MainCamera under /Main.

Search the project for PackedScene files in ui/.

Build a voxel block-world game with a player, blocks to place and destroy, and save slots.

A good first test that changes very little:

Using the Godot AI tools, add a DirectionalLight3D called "Sun" to the open scene, angle it 45 degrees down, then show me the scene tree. Don't touch any scripts.

What we tested

On October 18, 2026, we ran this on a Linux machine with no graphics card, using Godot 4.7.2 with software OpenGL and Godot AI v4.3.0. Enabling the plugin started its server and the dock showed "Server connected." Clicking Configure for Cursor wrote a config entry that pins godot-ai==4.3.0 via uvx. A test MCP client using that same command saw 47 tools. It created nodes in the open scene (including an imported .glb model), changed their scale and rotation, and read back the scene tree. The dock listed 26 supported clients.

Two things we noticed. First, the plugin does nothing if you run Godot in headless mode: the log says "plugin disabled in headless mode," so you need the normal editor window open. Second, it reports anonymous usage telemetry unless you set GODOT_AI_DISABLE_TELEMETRY=true before launching Godot (telemetry details). The README says it records an install ID, event, outcome, duration, platform and version, but not code, scene contents or file names.

Good to know (Godot)

C# is text-only. GDScript is fully supported. For C#, the plugin writes .cs files but doesn't build .NET or report compiler errors; build in the editor or run dotnet build (README).
It's local, with limits. Both hops are on your machine (ports 8000 and 9500 by default) and both are authenticated. The README says this doesn't protect you from a compromised program running under your own user account.
Alternatives (all Unofficial): Godot-MCP by IvanMurzak (Apache-2.0) and godot-mcp by hybridindie (PyPI godot-editor-mcp). We didn't test these.

Part 2: Roblox Studio

What you need

The latest Roblox Studio (Windows or macOS) and an MCP client (Roblox docs).

Roblox built the MCP server into Studio, so there's nothing extra to install. (Roblox's older standalone open-source MCP server on GitHub was marked "Product discontinued" on April 3, 2026, so skip tutorials that tell you to download it (GitHub).)

Step-by-step

1. Turn on the server. In Studio, open Assistant, click … → Manage MCP Servers, and switch on Enable Studio as MCP server (Roblox docs).

2. Connect with Quick Connect, if your client is listed. In Assistant Settings → MCP Servers, open the Quick connect dropdown and switch on your client. Quick Connect supports Antigravity, Codex CLI, Claude Code, Claude Desktop, Cursor, Gemini CLI and Visual Studio Code. If your client isn't listed, install it and restart Studio.

3. Or add it by hand. For other clients that use an mcp.json file, add this entry (Roblox docs).

Windows:

{
  "mcpServers": {
    "Roblox_Studio": {
      "command": "cmd.exe",
      "args": ["/c", "%LOCALAPPDATA%\\Roblox\\mcp.bat"]
    }
  }
}

macOS:

{
  "mcpServers": {
    "Roblox_Studio": {
      "command": "/Applications/RobloxStudio.app/Contents/MacOS/StudioMCP"
    }
  }
}

If you already have other servers, add the Roblox_Studio entry next to them and check your commas. Clients that want a command instead of JSON use cmd.exe /c %LOCALAPPDATA%\Roblox\mcp.bat (Windows) or /Applications/RobloxStudio.app/Contents/MacOS/StudioMCP (macOS). Restart the client afterwards.

4. Check it's connected. Go back to Assistant → … → Manage MCP Servers. A green indicator shows how many clients are connected.

5. Give the agent its first job. Studio's server lets the agent read and edit scripts, search the game tree, run Luau, start and stop playtests, read the output log, take viewport screenshots, simulate keyboard, mouse and character movement, search and insert Creator Store assets, and generate meshes, materials and procedural models (full tool list). Some prompts to try:

Look around first:

List the connected Roblox Studio instances, then describe what's in Workspace and ServerScriptService. Don't change anything.

A small script change with a test:

Create a server script in ServerScriptService that gives every player 100 coins in a leaderstats folder when they join. Then start a playtest, check the output log for errors, and stop the playtest.

Use the asset tools:

Generate a low-poly wooden treasure chest mesh and place it at the spawn point. Then take a screenshot of the viewport so I can see it.

Every tool call names a studio_id, so if you have several Studio windows open, the agent can call list_roblox_studios to pick the right one.

Good to know (Roblox)

Only connect clients you trust. Roblox warns that MCP clients can read and change content in your open places (Roblox docs).
Generation has limits. Roblox's Assistant documentation sets limits on its generation features. For example, procedural model generation is capped at 50 per rolling 24 hours (Roblox Assistant guide). The MCP page doesn't say whether the same caps apply when the MCP tools are called (not verified as of Oct 18, 2026).
Troubleshooting: restart both Studio and your client, check that the command path exists, and check your JSON for a missing comma or bracket.

Try it

Godot (Unofficial plugin):

Install uv, then download godot-ai-v4-plugin.zip (v4.3.0) from the Godot AI releases page and unzip it so you have addons/godot_ai/plugin.cfg.
In Godot 4.7+, enable it under Project → Project Settings → Plugins and wait for Server connected in the Godot AI dock.
Click Configure next to your client, restart the client, and ask: "Show me the current scene hierarchy."

Roblox Studio (built in):

In Studio, open Assistant → … → Manage MCP Servers and turn on Enable Studio as MCP server.
Turn on your client under Quick connect (Claude Code, Codex CLI, Cursor, Claude Desktop, Gemini CLI, VS Code or Antigravity) and restart it.
Look for the green indicator, then ask: "List the connected Roblox Studio instances, then describe what's in Workspace. Don't change anything."