Collaborate with AI to build your scenes in code, the language it’s actually fluent in. The AI works in code, you work in the editor, and two-way sync keeps both in agreement.
Not a wall of YAML. A flat, readable C# builder your AI can generate, validate, refactor, and sanity-check in one pass. A real compiler guarantees it’s a valid scene, and you can diff, review, and merge it like any other code.
public class MainMenu : ISceneDefinition
{
public void Build(SceneRoot scene)
{
var player = scene.Add("Player").Transform(pos: (0, 1, 0));
player.Component<Rigidbody>(rb => rb.Set(r => r.mass, 5f));
var door = scene.Add("Door").Transform(pos: (4, 0, 0));
scene.Add("OpenButton")
.Component<Button>(b => b.OnClick(door, nameof(Door.Open)));
}
}Drag Player somewhere in the Scene view and save. Your file rewrites itself to match, updating the right line in place.
player.Transform(pos: (0, 1, 0));
player.Transform(pos: (2.5f, 1, -3));↳ synced automatically from the editor
The construction is commodity. The value is the durable, living relationship between your code and your scene, and meeting AI in the medium it’s genuinely good at.
Edit the code or move something in the editor, and the other side updates to match, in place. It never wipes and rebuilds, so you can re-run it endlessly and every reference, prefab link, and object stays intact.
Stable identity across every re-run is the part a one-off build script can’t do. It’s the whole point.
LLMs are fluent in symbolic code. Hand the model the whole scene as one readable file it reads and edits in a single pass, not a stream of one-off editor commands.
On spatial-planning benchmarks top models score under 50%, yet ~95–100% on symbolic tasks.[4]
An MCP agent mutates the editor one action at a time and can stall mid-sequence, leaving a scene half-changed. CodeScenes applies a computed plan in place, all at once: it lands cleanly or not at all.
And because your scene is C#, invalid output won’t even compile.
For human-out-of-the-loop workflows, combine CodeScenes with the Unity CLI to build and test scenes autonomously. Nothing special connects them: the scene is just code, so the CLI drives it like any other project.
Your agent builds and tests it from the command line, then fixes its own scene.[7]
Your scene is clean C# in git: real pull-request diffs, line-by-line review, and git blame that means something.
You already know the pain of merging a .unity file. Code diffs like anything else.[1]
An MCP server holds dozens of tool schemas in context every session and round-trips every edit. CodeScenes puts one thing in context: the scene, as a file.
MCP tool schemas can eat the majority of a context window before you type a word. One measured setup hit 72%.[2]
Unity MCP servers are a serious, capable way to automate the whole editor. But for authoring scenes, making an AI drive the editor through a stream of granular tool calls fights the model instead of playing to it. We think scenes belong in code.
Context and reliability figures are general findings about MCP tool bloat and computer-use agents, not measurements of any specific Unity product. Comparison reflects our own view. See sources.
CodeScenes is early and I’d love your input on what to build. Reach out. I read everything.