Back to MCP overview

Summer Engine MCP guide

How to make games in LM Studio with Summer Engine

Use LM Studio as a free local MCP host: download a model, connect Summer Engine, and build playable games with no AI subscription. Summer Engine is the desktop game engine; LM Studio drives it through MCP with a model running on your own hardware.

LM Studio runs the model free on your hardware. Summer Engine shows the game.
01

Summer is free for MCP use

Download the desktop engine and connect it to your agent. The model runs on your own hardware, with no AI subscription and no per-message cost.

02

Open LM Studio

Run a free local model on your own GPU. The setup writes the MCP config so {agentLabel} can drive the engine directly.

03

Run a Summer game

The local model writes code. Summer Engine shows the Summer game, runs Play mode, and helps inspect errors. Export still requires the target templates and toolchains.

Install

Connect Summer to LM Studio

  1. 1

    Install Summer Engine

    Downloads the desktop engine app, about 1 GB. Skip it if Summer Engine is already installed, or if you only need assets and video for a game built in another engine.

  2. 2

    Sign in

    Opens your browser once. The CLI and the engine share the session.

  3. 3

    Connect LM Studio

    Writes the summer-engine MCP config and installs the Summer game-dev skills.

  4. 4

    Restart and build

    Open LM Studio, toggle on the summer-engine MCP server in the Program tab, and raise the loaded model's context length to 32k or higher.

Local setup notes

The 2026 specifics.

LM Studio is the whole stack

LM Studio is both the model runtime and an MCP host. The setup command writes ~/.lmstudio/mcp.json; toggle the summer-engine server on in the Program tab, and every engine call shows a confirmation dialog you can review before it runs.

Raise the context before you build

Models load with the context length baked into their metadata — often around 4k, which MCP tool schemas overflow silently. Set 32k or more when loading the model, and keep KV-cache quantization off for long agent sessions.

Models that hold up

gpt-oss-20b is the safe default on 12–16 GB of VRAM; Qwen3-Coder-30B and Gemma 4 27B are the 24 GB picks. On Apple Silicon, prefer the MLX builds — faster and more memory-efficient in long agentic loops.

Why this matters

LM Studio runs free on your hardware. Summer Engine shows the game.

A capable local model can operate a real engine through MCP. It opens scenes, places objects, runs the project and reads errors. Your GPU does the thinking, with no subscription and no per-token bill, and your prompts never leave your machine.

Start with LM Studio. Build inside Summer.

Install the desktop app for free, connect the MCP server, and make your first scene playable. No subscription required.