Stage Engine for AI: bring your own agent

A game project is normally opaque to a language model. It is scenes, binary resources, an editor that holds state, and a thousand conventions that live in a person's head. An agent asked to change something has to infer most of it and it gets things wrong in expensive ways.

Stage Engine exists to close that gap. It is a desktop app for Mac and Windows that bundles Godot, keeps the project in a real Godot layout, and exposes the work as things an agent can actually do rather than pixels it has to guess at. That is the layer, and it is why the code an agent writes here lands correctly more often than the same model writing into a bare project.

Director 1 is the agent that ships with the app. It writes GDScript into your project in response to what you ask for: behaviours, systems, rules, states. It does not rearrange your world, and it does not have a hidden channel into the builder. Code is its job.

The person using the app builds the world by hand and judges the result by playing it. That division is fixed and it is the design, not a stage on the way to something more automatic.

For most people this is the whole story. You never think about models, keys or protocols, and usage comes with your plan.

If you already work with your own AI tooling, Stage Engine can be driven by an external agent over MCP. Your agent, your model, your context — pointed at the app.

This is the path for people who have an existing setup they do not want to abandon, and for anyone who would rather keep the model choice in their own hands. It is not the default and nothing about the app requires it.

It is also the version that is honest about what an integration is. You are not getting a different Stage Engine; you are getting the same capabilities addressed by a different client.

The game runs locally. Godot ships inside the app and pressing play starts the game on your machine, at your framerate. There is no streaming, no remote render and no browser preview.

Building needs an internet connection. The agent, asset generation and world data all come over the network, so this is not an offline tool and there is no degraded offline mode. Playing what you built is a different matter and needs nothing.

Free to download for Mac and Windows, with a free plan. Indie is $10 a month, Pro is $50. Paid plans buy more usage rather than a different engine or extra features.

External agent access over MCP is part of what a paid plan covers. There is no seat pricing and no enterprise tier; this is a desktop app one person installs.

The interesting result is not that a model can write GDScript. It is how much less it has to invent when the surrounding project is arranged for it. Every ambiguity you remove — where a thing lives, what a scene is called, how a system is expected to be wired — is a decision the model no longer has to guess and get wrong.

That has a practical consequence. Reliability comes from the layer, not from reaching for a larger model, and a layer that makes a small model reliable is worth more than a prompt that makes a large one occasionally brilliant.

It is also why quality complaints here turn into work on the capability rather than a model upgrade. If the same request keeps producing the same wrong shape, the honest fix is upstream of the model.