KKitwrightGitHub
v0.6.0 · Unity 2022.3+ · MIT

Your AI assistant, inside the Unity Editor

An MIT-licensed MCP server that lets Claude Code, Cursor, Codex and any other MCP client operate directly inside your running Unity project.

150
built-in tools
35
modules
32
tools in core profile
MIT
license
Clients

Works with the client you already use.

Vendor agnostic. Anything that speaks MCP can drive your Editor.

Claude CodeCursorLM StudioWindsurfCodexVS Code CopilotKiroTrae
How it works

Describe the change. It happens in the Editor.

execute_code first
01 / FEATURE

execute_code first

One in-memory C# execution tool instead of fifty narrow ones. Roslyn compiles the snippet, it runs on the editor thread, and the agent gets the full Editor and runtime API surface — no .cs files written under Assets/, no domain reload.

Play mode automation
02 / FEATURE

Play mode automation

Enter play mode, simulate keyboard and mouse input, capture the game view, read the console, and validate behavior — all from the same MCP session.

Auto-Undo and structured logs
03 / FEATURE

Auto-Undo and structured logs

Implement IGameWrightCommand and every created, modified or destroyed object joins editor Undo. The response carries logs, created, modified, destroyed and returnValue, so the agent verifies changes without re-querying the scene.

Focused by default
04 / FEATURE

Focused by default

The core profile exposes 32 high-signal tools to keep the client's tool list clean. Switch to full for all 150 when you need them, and edit exactly which tools each profile exposes.

The recommended execute_code template:
CommandScript.cs
public class CommandScript : IGameWrightCommand
{
    public void Execute(ExecutionContext ctx)
    {
        var go = GameObject.CreatePrimitive(PrimitiveType.Cube);
        ctx.RegisterObjectCreation(go);   // auto-Undo + tracked
        ctx.Log("Created {0}", go.name);
    }
}
Details

The parts that matter once you use it daily.

Structured returns

Every tool returns {success, message, data} with stable instanceId fields, so agents chain by_id calls instead of re-resolving by name.

Safety guards on by default

The filesystem guard blocks destructive snippets, broad System.IO writes and traversal paths. Not a full sandbox — clients can override per call.

Resources and prompts

Live project context, scene and selection state, compile errors, interaction history, plus workflow prompts like fix_compile_errors.

One-click client config

Generate MCP config entries for Claude Code, Cursor, LM Studio, VS Code, Codex and more, straight from the Unity window.

Editor-only

No runtime components are added to your built game. The server listens on 127.0.0.1:8765 and nothing leaves your machine.

Built-in updating

GameWright > Check for Updates refreshes Git installs in place or imports the latest unitypackage automatically.

Compare

Against Unity's own AI Assistant.

Unity AI Assistant v2.7.0-pre.2, as documented in May 2026.

AreaGameWright MCPUnity AI Assistant
Minimum Unity version2022.36000.3 (Unity 6 only)
LicenseMIT, open sourceProprietary, Unity ToS
DeploymentLocal HTTP server in EditorEditor + relay + Unity Cloud
BillingFree, bring your own clientCredits via Unity Dashboard
Tool exposure150 tools, core (32) / full~15 MCP tools
Play mode validationEnter, input, capture, logs, exitEnter and exit only
Offline tool callsYesNo
FAQ

Questions worth answering.

Is it really free?

Yes. MIT licensed, no credits, no account. You bring your own MCP client and your own model.

Which Unity versions work?

Unity 2022.3 and newer, including Unity 6. The package is Editor-only and adds nothing to your build.

Does my project code leave my machine?

The MCP server itself is local, on 127.0.0.1:8765. Whatever your AI client sends to its own model provider is governed by that client.

Do I need Python?

No. The Unity-side plugin is a single Unity package with no external daemon and no Python requirement.

Is execute_code sandboxed?

There are default-on safety checks and a stricter filesystem guard, but it is not a complete sandbox. Treat it as you would any code your agent runs locally.

Can I add my own tools?

Yes. Tools are discovered by attribute, so a custom tool is a class and an attribute away.

One Git URL and you are connected.

No account, no credits, no Python. Add the package and start the server.