Unity MCP: agents that build scenes, write scripts and run tests in Unity

Writing a Unity game with Cursor or Claude Code gets awkward because the agent can’t see the editor. Did the script attach? How many objects are in the scene now? Did the tests pass? You end up relaying screenshots. Unity MCP, also branded AI Game Developer, installs a plugin in the editor that opens the asset database, scene hierarchy, components and test runner to the agent. It had about 4.4k stars as of 2026-10-06, among the most for any Unity MCP project.

What it does

• Assets and packages: find, create and move assets, create materials, edit prefabs, and install UPM packages with package-add.
• Scenes and objects: gameobject-create, gameobject-component-add and gameobject-modify, plus screenshots from a camera, the Game view or the Scene view so the model can check its work.
• Scripts and editor: script-update-or-create writes C#, script-execute compiles and runs code through Roslyn, reflection-method-call invokes any method, tests-run runs EditMode and PlayMode tests, and console-get-logs reads the log.
• Profiling: toggle the Profiler, capture frame timing, read memory and rendering stats.
• In-game runtime and custom tools: the plugin can ship inside a built game to drive NPCs or debug live, and one attribute in your code turns a C# method into a tool.

Who it’s for

• Indie developers who want an agent to block out levels, tweak component values, run the tests and report back.
• Technical artists and tools programmers who expose their project’s own workflows as custom tools.

Setup

Needs the Unity Editor. Add the plugin through OpenUPM:

openupm add com.ivanmurzak.unity.mcp

Then open Window/AI Game Developer in Unity and click Configure to write the MCP config, or Auto-generate to create Skills. The .unitypackage installer and the unity-mcp-cli tool do the same job.

Our take

There are several Unity MCP options. This one stands out for a broad tool set, frequent releases and an in-game runtime mode that’s rare among its peers. Know how it connects first: the CLI quick start signs you in to ai-game.dev and by default points your client at its hosted endpoint. To keep a third party out of the loop, use the local binary under the project’s Library/mcp-server/ over stdio, as the README’s table shows. script-execute and reflection calls let the model run arbitrary code in your editor, and assets-delete can remove files, so keep the project under version control. Licensed Apache-2.0.