Skip to content

Making games with local LLMs

By Eduardo Orellana · Updated

A model you host yourself can drive Flockbay, because the connection is MCP and MCP does not care where the model lives. What it cannot do is remove the network from the rest of the job. Free AI to start: Flockbay's AI and your own Claude, ChatGPT, agent or API key are free for your first build; after that, a plan keeps you building. It runs on Windows, macOS and Linux, on the official Godot engine.

The main page on this subject is Connect your own agent.

AppImage, x86_64. Ubuntu 22.04, Debian 12, Fedora 39 or newer.

Intel or AMD Chromebook, with Linux turned on.

Download for macOSDownload for WindowsDownload for Linux

Three reasons people want this, answered separately

Cost. A model you host has no per-token bill. Connecting your own agent over MCP is free for your first build, and a plan keeps it building after that.

Privacy. What your agent thinks, and the text you type at it, stays on your machine if the model is on your machine. That part is real. What is not private is everything else the app does, which is covered in the next section and is the reason this page exists.

Control. This one holds up best. Your model, your version, no deprecation notice, no rate limit, no change in behaviour on a Tuesday because someone shipped an update. If you have built habits around a specific local setup, keeping it is a legitimate reason to bring your own agent rather than use Flockbay.

More on this: Flockbay vs Summer Engine: a free alternative on the official Godot

What still leaves your machine

Flockbay is online only for building. Asset generation runs over the network, and so does the Flockbay account the connection needs. Neither of those is the language model, and neither becomes local because your model did.

Building needs an internet connection, with a local model too. If your requirement is a genuinely air-gapped game-making setup, Flockbay is not it and no configuration makes it it. Better to hear that here than three hours in.

The one thing that does run entirely on your machine is the game. Godot ships inside the app, and when you press play it runs locally with your machine's memory and framerate. No streaming, no browser preview, no render queue. A finished game does not need the connection at all.

Why the editor and the running game matter more for a small model

Most local models are smaller than the hosted frontier ones, and the usual advice is that they therefore cannot do agentic work on a complex codebase. That advice is broadly right when the codebase is a raw game project, because a raw game project asks the model to invent almost all of its own context.

Flockbay hands it the Godot editor and the running game instead. A node added through the editor is written by the editor, UIDs and undo step included, rather than as lines of .tscn text the model has to get right. A change is checked by playing the game and reading stdout, the debugger and values off the live scene tree. Every check the tools run is a mistake the model can see and fix instead of guess at.

The honest version of the claim: a local model gets the same tools as any other, which is what makes a self-hosted model worth attempting at all. We have not benchmarked local models here and will not tell you which one to run or which is best. Try yours.

How the connection works, without the configuration

Flockbay ships one MCP server inside the desktop app. It carries the scene tools, the asset library, and the game itself, which it can edit in the open Godot editor, run, look at and play. An MCP-capable agent on the same machine can call it. That agent can be backed by whatever model you like, including one you are serving locally.

We have not built anything specific to any local-model runner. The Flockbay plugin covers Claude Code, Codex and Gemini CLI on Windows, and every other client connects to the same server directly. Popular MCP clients should work, behaviour varies between them, and we have not tested every one.

We also do not publish the setup steps here. Command names, flags and config file contents change, and a stale instruction on a page like this is worse than sending you to the app, where the current details live.

More on this: What is MCP for game development?

The part no model does

You judge the game. The model builds the scenes, whether it is local, hosted or Flockbay's own, and can play the game to check its work. Deciding that it looks right is yours, and when one call is pure taste, make it yourself in the Godot editor's own 3D or 2D view.

And you judge the game. You play it, decide whether it is right, and say what is wrong. Handing work back is the normal loop here. Not writing the code has never implied not having an opinion about the result.

If none of the three reasons at the top of this page apply strongly to you, use Flockbay's own AI. It is in the app, it agrees a brief with you before it builds, it already knows Flockbay's tools, and there is nothing to connect or maintain.

More on this: AI agents and game development

A local model needs an agent in front of it

A local model server, Ollama or LM Studio, runs the model; the agent is what calls tools. Give the connection to an agent that speaks MCP and runs your local model: Goose, OpenCode, Cline, or LM Studio’s own chat. Goose is shown here.

The Flockbay app writes this for you with the real paths filled in: open Connectors, press Set up on Connect your agent (MCP), press Turn on, pick Goose and press Copy. The three paths are specific to your machine and stay the same from one sitting to the next, so you paste it once. Below is the shape, with the paths as placeholders.

Add it to ~/.config/goose/config.yaml, under the extensions you already have, and point Goose at your local model as usual.

More on this: Making games from Goose

YAML
extensions:
  flockbay:
    name: flockbay
    type: stdio
    cmd: '<node>'
    args:
      - '<coordinator>'
      - '--external-stage-mcp'
    envs:
      FLOCKBAY_MCP_DESCRIPTOR: '<descriptor>'
    enabled: true
    timeout: 300

Questions

Can I use Flockbay offline with a local model?
No. Asset generation and your Flockbay account go over the network, so building needs a connection whatever the model is doing. The finished game runs locally without one.
Which local model works best?
We have not measured local models. Every model gets the same tools, including running the game and reading its errors, so try the one your machine runs well.
Which MCP clients are supported?
Any client that speaks MCP should be able to reach the server. We have not tested every one and behaviour will vary.
Is bringing my own model free?
Yes, for your first build: connecting is free, and a plan keeps it building after that. You pay your agent's provider for the model.
Does my model have to write the scenes as well as the code?
No. The scenes are not only text your model writes. Your agent builds them by calling Flockbay's scene tools, which make real nodes in the scene tree through the Godot editor, and the editor is open beside it if you want to change something yourself.
What runs the game itself?
Godot, bundled inside the desktop app for Windows, macOS and Linux. The project is a real Godot project on your disk.

Free to download for Windows, macOS and Linux. Your game runs locally on the bundled Godot.

Download Flockbay free