Capybara Version 2 is available. We recommend all new users start on Version 2 — Version 1 support will be deprecated over time.

Visit Version 2 →

Capybara

Game engine

2026

DocsPricingDiscordGitHubSign in →

# Game engine for coding agents

An open source game engine for coding agents, plus an MCP asset pipeline that turns AI images into maps, characters, VFX, HUDs, and more—ready to drop into a playable game.

Get an API keyRead the docs →
$ claude mcp add capybara-mcp
  -e CAPYBARA_API_KEY=cb_live_…
  -- npx -y capybara-mcp@latest

## Examples

A walkthrough of building with a coding agent and the Capybara MCP, plus one-shot intros using GPT-5.5 — prompt in, playable game out.

01 · Engine walkthroughCoding agent + MCP
02 · One-shot example 1Built with GPT-5.5~$2.18 per game
03 · One-shot example 2Built with GPT-5.5~$1.90 per game

## Why this exists

Coding agents are already good at scaffolding gameplay. Assets are where things slow down. Most engines expect hand-made tilesets and sliced sprites. AI images arrive as whole scenes, with light and shadow already in place.

Capybara is built around that. The engine and asset pipeline share the same metadata language, so a generated image can be placed, collided with, and iterated on as a real game piece.

It took about eleven months: an open source game engine for coding agents, and an asset pipeline over MCP so the agent can ask for what it needs while it works.

## Built for agents

Interfaces stay simple, and dependencies stay light. The loop is gentle: call a tool, receive structured assets and metadata, write plain game code, play the preview.

Today we focus on 2D top-down at a Stardew-like angle, with two sideways facings. Four-way is next, and we will ask on Discord which camera angles to support after that.

## The asset MCP

The asset pipeline is the commercial piece. Pricing follows API and GPU cost—each request does real work to turn an image into something playable. You can pass a style reference for a base map, and that look carries through characters, props, and the rest of the set.

What you can generate:

  • Maps — sections of a larger world, with bounding boxes, depth masks, and the metadata needed to work out of the box.
  • Characters — matching the map’s art language, with variants for clothes, tools, and other loadouts.
  • VFX — layered effects that sit on top of maps and bring scenes to life.
  • HUDs — each one includes an HTML overlay with coordinates and placeholders, so the agent can wire clicks and state.
  • Multi-props — groups that stay consistent across frames, like crop lifecycles.
  • Prop overlays — quiet state changes without full animation, like an open door over a closed one.
  • Map extensions — still in beta. Grow a world with matching terrain and detail.
  • Background music — tracks that fit the scene you are building.

## Server features

Optional server pieces help you stay on gameplay: guest and passwordless signup for players, data storage, and multiplayer. That backend will be open sourced when it is ready. Until then, the SDK infra is free to use.

The engine runs fine on its own. The SDK is there when you want auth and persistence handled, so more of the agent’s attention can go to design and systems.

## Getting started

Sign in, add credits, mint an API key, then paste the setup command into Claude Code, Codex, or another MCP client. Describe a scene, review what comes back, play the preview, ask for the next room.

The engine is open source on GitHub. The asset MCP is there when you need generation. Shared metadata keeps the two sides feeling like one tool.

Open the developer console →View the engine →