Making games from Goose

The way Goose is put together makes this straightforward to picture. You add capabilities by adding servers, and the agent works out what it can call. Stage Engine ships two of them inside the desktop app: one exposing the Stage capability layer, one exposing assets.

That means the game engine does not have to be a separate world you switch into. It can sit beside whatever else you have connected, and an agent can move between them in a single piece of work. Pull a reference from somewhere, put something into the project, come back. Nothing about the protocol prevents that.

We should say what this is not. There is no Goose extension we publish, no listing we maintain, and no partnership. We speak MCP, Goose speaks MCP. Any client that speaks it should work, and we have not tested every one, so expect behaviour to differ.

Goose lets you swap the model underneath, which makes it a good place to notice something we care about a lot. The quality of what an agent produces against a game engine depends less on the model than people assume, and much more on what the model is being asked to invent.

Point an agent at a bare Godot project and it must invent almost everything: how a scene is meant to be structured, what belongs where, how terrain and water and weather relate. That is where a stronger model appears to help, because it is guessing more skilfully. Guessing skilfully is still guessing.

The capability layer removes the guess. Operations are stated at the level a game is described in, and each one is code that already works. Composing against that is a smaller job, and a smaller job is one a cheap model can do reliably. We are not going to tell you which model to run. We will say that the layer is doing more of the work than the choice is.

It is worth tempering the appeal of a single agent that can reach everything. Connecting more servers gives an agent more to consider on every turn, and the practical effect is often a slower, less decisive agent rather than a more capable one. If a session is about the game, a narrow set of tools tends to go better.

The other limit is the one that does not move. The world is built by hand, in the builder inside the app: terrain, water, weather, buildings and characters, placed by you and looked at by you. That is the part where taste is the input and where a sentence is a worse instrument than your own hands. It is the design rather than a gap in the protocol.

And you are the one who says whether the game is right. You play it, you find what is wrong, you go back in. That loop is the normal shape of the work here.

The Stage Engine download is free and there is a free plan, which includes Director 1 inside the app. External MCP access, which is what this page is about, is part of the paid plans: Indie at $10 a month, Pro at $50. Those buy usage and that access, not a different engine.

Stage Engine is a desktop application for Mac and Windows. It is downloaded and installed, with its own copy of Godot inside it, so there is no separate engine to set up and no browser version. The project is a real Godot project in a folder on your machine, which is why an agent can work on it at all.

Building needs an internet connection, because the AI, asset generation and world data come over the network. That is true whether the model is running on your machine or somewhere else. Playing a game you already built does not need it.

For connecting a client, follow the app and its documentation. We do not print configuration examples here on purpose: they drift between versions, and a wrong one costs you more than an absent one.