esengine /estella
A fast 2D and 3D game engine — TypeScript SDK, C++/WebAssembly core, visual editor. Ship one project to web, desktop, WeChat MiniGames, playable ads, and native Android / iOS.
★ 939⑂ 195👁 37⑃ 123● 87▤ 127
939
stars
195
forks
37
watchers
123
prs
87
issues
127
releases
README
Estella A 2D and 3D game engine whose editor is an MCP server. Your AI agent builds the game with the same 167 tools the editor's own UI uses — then ships it to the web, mini-game platforms, playable ads, desktop and native mobile. Website · Agent · Getting Started · Documentation · QQ群 481923584 Every frame is one MCP tool call into the editor, captured as it happened — no mouse involved. Your agent drives the editor The Estella editor speaks the Model Context Protocol. Connect Claude Code, Cursor, Codex or any MCP client, and the agent works inside the real editor rather than around it: - It builds — creates projects, spawns entities from the editor's own Create catalog, writes component fields, makes prefabs, paints tilemaps, writes and type-checks scripts. - It looks — takes screenshots of the viewport and the running game, replays a frame draw by draw, compares the game with reference images, and audits what a scene's look lacks. - It plays — enters Play mode, steps time deterministically, sends input and clicks UI by name, runs scripted playtests, then reads the running world back. - It makes assets — images, sound effects, music and 3D models, each answered with a picture of what it made. - It ships — reads project health, exports the game, and drives a build running on a connected phone. Every call goes through the editor's command surface: each one is one undo step, you watch it happen in the window, and you can take the mouse back at any time. The editor starts read-only; an agent edits only after you allow it in Window → Connect AI Tools…, which also writes the client's MCP config for you. | Group | What the agent gets | |---|---| | core | scene tree, entities, components, play / step / input, screenshots, logs, undo | | assets · prefab · authoring | asset documents and import settings, tilemaps, prefabs and variants, physics bodies, UI events | | look · render | reference images beside the game, light and field sweeps, frame capture and replay, shadows, LOD, bakes | | code · project | scripts, symbol search, type-check, project settings, health, export | | image · sound · modeling | draw sprites and textures, compose sound and music, build 3D models from steps | | play · view · device | gamepads, time scale, scripted playtests, editor view, and a game running on a connected phone | → Connecting an agent Why Estella, when Unity and Unreal have MCP too? An MCP server can be bolted onto any editor. What an agent can actually get done depends on what sits behind it. - The protocol is the editor, not a plugin on it. The 167 tools call the pipelines the editor's own menus, inspector and toolbar run, so an agent's edit is the edit a person would have made, and undoes the same way. Calls are schema-checked and permission-gated by the editor itself. - The agent can check its own work. It steps a paused game frame by frame, so "play three
Latest release
v0.106.0 v0.106.0 2026-11-15Added
- A native build's sockets connect to wss://. createSocket({ url: 'wss://…' }) now opens on the desktop, iOS and Android runtimes, where it used to report that only ws:// is supported and close. TLS is the operating system's, as it already is for fetch: Network.framework on macOS and iOS, Schannel on Windows, libcurl on Linux and the platform's SSLSocket on Android, under the same WebSocket framing a ws:// connection uses. The certificate chain and the host name are checked against the system's trust store; a connection that fails either reports the reason as an error event and closes w