Making games from OpenCode

It is worth being blunt about this before anything else. We have not built anything specific to OpenCode. There is no extension, no listing, no plugin, and nobody here has certified a version of it. What exists is that Stage Engine runs two MCP servers inside the desktop app, and OpenCode is a client that speaks MCP.

That is a weaker claim than a partnership and a more durable one. Nothing about the connection depends on us tracking someone else's release notes. Any client that speaks the protocol should be able to reach the same surface, and OpenCode happens to be one of the popular ones.

It also means behaviour will vary. Different clients handle long tool listings, errors and permissions differently, and we have not tested every one. The app and its own docs are the place to get the current connection details from; anything written here about flags or file contents would go stale and be worse than nothing.

Two servers, with different jobs. One exposes the Stage capability layer, which is the set of operations that describe a game rather than a file. The other exposes assets. Between them, an agent has something to call that is closer to the level a game is discussed at.

The contrast is with pointing a terminal agent at a bare Godot project directory, which many people have tried. What the agent finds there is a folder of scenes, resources and scripts, with the meaning of the game distributed across all of them and written down nowhere. It can read and edit, competently, and it will still be inferring what a scene is for on every single turn.

The capability layer is the difference between an agent editing files that happen to make a game and an agent operating a game. That is the part that took the work, and it is the part that is worth pointing OpenCode at.

The world is built by hand. Terrain, water, weather, buildings, characters and where they stand are placed by a person in the builder, in the app, with a mouse. That is not a limitation we intend to remove later. It is the design, and it holds whether the code is coming from Director 1 or from your own agent in a terminal.

So a session is two windows, not one. The app is open in front of you and you are building in it. OpenCode is beside it, writing the behaviour. If you were expecting a workflow where you never leave the terminal and a world assembles itself, that is not this, and no amount of protocol makes it that.

The upside of the split is that the two halves fail separately. When the world looks wrong you fix the world, directly, in seconds. When the behaviour is wrong you say so to the agent. Neither problem hides inside the other, which is the usual reason a long agent session goes sideways.

Director 1 ships in the app and writes GDScript into the project. It is aimed at the capability layer already, so there is nothing to connect and nothing to keep working. For most people it is simply the shorter path and it is the one we would suggest first.

Bringing OpenCode is worth it when you have real reasons: you want your own model choice, you already have a working set of habits in the terminal, or you are doing something around the game as well as inside it and would rather one agent held the whole thing. Those are good reasons. "I have a terminal open anyway" is a weaker one.

You can also do both. Nothing forces a choice, and using Director 1 for most of the work while driving the app from OpenCode for the parts you care about is a perfectly reasonable way to work.

Stage Engine is a desktop application for Mac and Windows, and Godot ships inside it, so there is no separate engine install. The project it operates on is a real Godot project in a folder on your machine. When you press play the game runs locally on that bundled Godot.

External agent access over MCP is part of the paid plans. Free is $0 a month, Indie is $10 and Pro is $50. What those buy is usage and this external access, not a different engine and not different capabilities.

Building needs an internet connection, because asset generation and world data come over the network. That stays true no matter where your agent or your model is running. Playing what you have built does not need the connection.