An AI agent that works inside a Godot project

The difference between an assistant and an agent is not the model. It is the surface. An assistant answers about your project. An agent acts on it — it can look at the world, change files, and see whether the change worked.

For a game, that surface has to include the world itself. Reading the scripts is not enough, because most of what a game is lives in the scene: what is placed where, how the ground is shaped or which cells of it are ground at all, what a character is standing on. An agent that can only read text is reconstructing the world from a shadow of it.

Flockbay works against a world it can actually address, and which world that is depends on the answer given when the project was made: 2D or 3D, shown as two pictures, with no default. In a 3D project it is terrain, water, weather, buildings and characters. In a flat one it is the tile ground and the coverage the tiles are derived from, the band of scenery on the horizon, every sprite placed in the scene, and a character body whose feel is a set of named numbers rather than a script to be rewritten. Either way these are things it can reason about directly rather than infer from a folder of scripts, and that is the difference between an agent that composes reliably and one that produces confident nonsense.

More on this: Does AI game making actually work?

Flockbay writes code. You describe behaviour — the door needs a key, the wolves hunt at night, falling in the river costs you a life — and it writes the GDScript into your project. The code is visible. You are not asked to read it, and nothing stops you.

What it does not do is build the world for you. That is you, by hand, in the builder — shaping terrain and outlining water in a 3D project, dragging tile ground in and back out again in a flat one, and putting things down yourself either way. This is a deliberate division rather than a missing feature. A sentence is a bad instrument for a hillside, and taste does not compress into a prompt.

It also does not decide whether the game is good. You play it and say. That handoff is the normal loop, not a failure of the agent.

Flockbay is a desktop app for Windows today, with a Mac release later, and Godot inside it. The agent is operating on files on your disk, in a real Godot project, and when you press play the game runs on your machine at your framerate. Nothing is streamed and nothing is rendered elsewhere.

That matters for an agent specifically, because the feedback loop is short. It writes, you run it locally, you see it fail, you say so. A round trip through someone else's render queue would make that loop slow enough to stop being a loop.

Building does need an internet connection, since the chat, asset generation and world data all come over the network. Playing the result does not.

More on this: What is MCP for game development?

Flockbay exposes its capabilities over MCP, which means you are not required to use Flockbay as the only agent in the loop. If you already work with an agent you like, it can drive the same surface. Connecting your own agent over MCP is free.

The point of that is not novelty. It is that the valuable part is the capability layer — the fact that the world is legible and operable at all — rather than any particular model. If the layer is right, a cheaper model composes reliably against it, and that is the design being aimed at.

The download is free and the builder is free to use. Flockbay is free on your own AI: your Claude or ChatGPT plan, your own agent, or an API key.

More on this: Game development over MCP