Unity AI Bridge

io.github.butterlatte-zhang由 GitHub 账号 butterlatte-zhang 发布v1.0.0更新于 Oct 24, 2026

Remote-control Unity Editor from any AI IDE via file-based IPC. 62 tools, zero dependencies.

已验证STDIO仅桌面开发者工具AI 与机器学习

概览

AI 生成的概览

让 AI 助手在本地远程控制 Unity 编辑器:查看场景、编辑资源与脚本、运行测试、性能分析,并驱动播放模式测试。

可以做什么
Unity AI Bridge 提供约 65 个 Unity 编辑器工具,涵盖 Scene、GameObject、Assets、Prefab、Script、Editor、Reflection、Screenshot、Runtime、Console、Profiler、Package、Light Probe 和 Tests 等类别。助手可以打开和保存场景、创建或修改游戏对象与组件、读写脚本、截图、读取控制台日志、采集性能分析快照,并在播放模式下查询或调用运行时状态。它与 Unity 插件通过基于文件的 IPC 通信,而不是网络套接字,并同时支持 skill/CLI 模式和 MCP 模式。
适用场景
当助手需要深入 Unity 项目内部而不只是编辑文件时使用:查看或修改场景与预制体、检查编译错误、运行 Unity 测试、做性能分析,或执行自动化的“编写—测试—修复”播放测试循环。适用于 Windows 或 macOS 上的 Unity 2022.3 LTS 及更新版本项目。
运行要求
本地 Unity 项目需安装 com.aibridge.unity 包(通过 git URL 或复制包目录),并配置支持 skill 或 MCP 模式的 AI IDE。该桥接是纯 Python 标准库实现的 CLI/MCP 服务器,无需 pip、npm 或 Node.js 运行时。未声明任何账号、API 密钥、环境变量或请求头。
安装前请注意
这些工具会写入和删除项目内容:脚本、资源、预制体、文件夹和游戏对象都可能被创建、修改、移动或删除,编辑器状态、播放模式和包安装也可能被更改。反射和运行时调用可以调用正在运行项目中的方法。它完全在本地运行且不开放网络监听,但文件 IPC 通道位于用户临时目录中,任何能访问该目录的程序都可能与之交互。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Unity AI Bridge,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

English | 中文

Unity AI Bridge

Remote-control the Unity Editor from any AI IDE — no ports, no dependencies, just works.

[Unity 2022.3+] [License: Apache 2.0] [GitHub stars] [Release] [Glama]

https://github.com/user-attachments/assets/4e8b3f85-b209-406f-a96e-f8b8eddc9160


Why Unity AI Bridge?

Most AI coding assistants can read and write files, but they are blind to the Unity Editor — they can't inspect your scene, tweak materials, run tests, or profile performance. Unity AI Bridge gives AI full editor access.

Key Advantages

  • 65 tools, 15 categories — Scene, GameObject, Assets, Prefab, Script, Profiler, LightProbe, Screenshot, Runtime, Tests, and more. Covers the full editor workflow, not just file I/O.
  • File-based IPC, not WebSocket — No open ports, no firewall issues, no connection drops. Survives recompilation, play-mode transitions, and editor restarts gracefully.
  • Zero external dependencies — Pure Python stdlib CLI/MCP server, self-contained C# Unity package. No pip install, no npm, no Node.js runtime.
  • Every major AI IDE — Claude Code (Skill mode), Cursor, GitHub Copilot, Windsurf, Claude Desktop (MCP mode). One Unity plugin, all IDEs.
  • 5-line extensibility — Add custom tools with [BridgeTool] attribute. Auto-discovered, auto-serialized, auto-documented. No registration code needed.
  • Production-tested — Built for and battle-tested in a large-scale open-world Unity game (50+ developers, 2M+ lines of C#).

vs Unity 6 AI Gateway

Unity 6.2 introduced an official AI Gateway with MCP support. Both projects share the same goal — giving AI agents editor access via MCP — but differ in important ways:

Unity AI BridgeUnity 6 AI Gateway
Unity version2022.3 LTS+6.2+ only
Tool coverage65 tools across 15 categoriesGeneral-purpose (Scene, Assets, Script, Console)
Deep toolingProfiler (snapshot, hotpath, stream), LightProbe, Reflection, Package ManagerNot yet available
IPC mechanismFile polling (~100ms)Unix Socket / Named Pipe
Extensibility[BridgeTool] attribute — 5 linesTBD

In practice, the ~100ms file-polling latency is imperceptible because AI agent think-time dominates each round trip. File IPC also makes cross-process debugging trivial — just inspect the JSON files on disk.


Quick Start

AI-native project — Copy the prompt below and send it to your AI coding assistant. The setup guide is written for AI to follow — you don't need to run any commands yourself.

Help me install Unity AI Bridge by following this guide:https://github.com/butterlatte-zhang/unity-ai-bridge/blob/main/docs/SETUP.md

If you prefer manual setup:

  1. Unity Package — In Unity: Window > Package Manager > + > Add package from git URL:

    https://github.com/butterlatte-zhang/unity-ai-bridge.git?path=Packages/com.aibridge.unity

    Or manually copy Packages/com.aibridge.unity from this repo into your project's Packages/ directory.

  2. IDE Integration — Copy .claude/ to your project root, then configure your IDE per docs/SETUP.md.

Supports: Claude Code (Skill mode), Cursor, GitHub Copilot, Windsurf, Claude Desktop (MCP mode).


Tool Categories

65 tools organized into 15 categories:

CategoryCountTools
Scene7scene-open, scene-save, scene-create, scene-list-opened, scene-get-data, scene-set-active, scene-unload
GameObject11gameobject-find, gameobject-create, gameobject-destroy, gameobject-modify, gameobject-duplicate, gameobject-set-parent, gameobject-component-add, gameobject-component-destroy, gameobject-component-get, gameobject-component-list-all, gameobject-component-modify
Assets11assets-find, assets-find-built-in, assets-get-data, assets-modify, assets-move, assets-copy, assets-delete, assets-create-folder, assets-refresh, assets-material-create, assets-shader-list-all
Prefab5assets-prefab-create, assets-prefab-open, assets-prefab-save, assets-prefab-close, assets-prefab-instantiate
Script4script-read, script-update-or-create, script-delete, script-execute
Object2object-get-data, object-modify
Editor4editor-application-get-state, editor-application-set-state, editor-selection-get, editor-selection-set
Reflection2reflection-method-find, reflection-method-call
Screenshot1screenshot-capture
Runtime2runtime-query, runtime-invoke
Console1console-get-logs
Profiler5profiler-snapshot, profiler-stream, profiler-frame-hierarchy, profiler-hotpath, profiler-gc-alloc
Package4package-list, package-search, package-add, package-remove
Light Probe5lightprobe-generate-grid, lightprobe-analyze, lightprobe-bake, lightprobe-clear, lightprobe-configure-lights
Tests1tests-run

Architecture

┌──────────────────────────────────────────────────┐│                   AI IDE                         ││  (Claude Code / Cursor / Copilot / Windsurf)     │└──────────┬────────────────────┬──────────────────┘           │                    │     Skill mode            MCP mode           │                    │           ▼                    ▼    ┌─────────────┐    ┌──────────────┐    │  bridge.py  │    │ mcp_server.py│    │  (Python)   │    │  (Python)    │    └──────┬──────┘    └──────┬───────┘           │                  │           └────────┬─────────┘                    │              File-based IPC            (request / response)                    │                    ▼    ┌───────────────────────────────┐    │     Unity Editor Plugin      │    │   (com.aibridge.unity)       │    │                              │    │  BridgePlugin ← polls files  │    │  BridgeToolRegistry          │    │  BridgeToolRunner            │    │  [BridgeTool] methods        │    └───────────────────────────────┘

Dual-channel design: The same Unity plugin serves both Skill mode (direct CLI) and MCP mode (protocol server). Both channels communicate through the same file-based IPC — a pair of request/response files on disk. No network sockets, no port conflicts, no firewall rules.

Why file IPC? Unity's main thread is single-threaded and blocks during domain reload. File polling is the most reliable way to survive recompilation, play-mode transitions, and Editor restarts without losing messages.


Beyond Editing — AI as Game Tester

Most Unity AI tools stop at file editing. Unity AI Bridge goes further — it turns Claude Code (or any AI IDE) into a game testing harness.

CapabilityTraditional AIWith Unity AI Bridge
Write C# code:white_check_mark::white_check_mark:
Check compilation errors:x::white_check_mark: console-get-logs
Enter / exit Play Mode:x::white_check_mark: editor-application-set-state
Trigger game actions:x::white_check_mark: runtime-invoke
Read runtime game state:x::white_check_mark: runtime-query
Take screenshots:x::white_check_mark: screenshot-capture
Full closed loop: Write → Test → Fix → Repeat:x::white_check_mark:

AI Playtest Loop

AI writes code → compiles → enters Play Mode → observes state → judges → fixes → repeats     ↑                                                                         │     └─────────────────── fully automated loop ────────────────────────────────┘

The pattern: Act → Wait → Observe → Judge → Repeat

  • Act: runtime-invoke calls static methods to trigger game actions
  • Wait: wait_playmode.py / wait_compile.py handle timing
  • Observe: runtime-query reads MonoBehaviour fields + screenshot-capture for visuals
  • Judge: AI analyzes state/screenshots to decide PASS/FAIL

Auto-Playtest Example | AI Closed-Loop Guide | Playtest Tool Reference


Add Your Own Tools

Expose any static method to AI with a single attribute:

csharp
using UnityAiBridge;
[BridgeToolType]public static partial class CustomTools{    [BridgeTool("custom-greet")]    [System.ComponentModel.Description("Say hello")]    public static string Greet(string name = "World")    {        return $"Hello, {name}!";    }}

The bridge discovers tools at Editor startup via reflection. No registration code, no config files. Parameters are automatically mapped to JSON Schema for the AI to call.


Security

Unity AI Bridge runs entirely on your local machine. The file IPC channel is scoped to your user's temp directory, and no network listeners are opened.

See SECURITY.md for details.


Compatibility

Unity VersionRender PipelineStatus
2022.3 LTS+Built-inSupported
2022.3 LTS+URPSupported
2022.3 LTS+HDRPSupported
6000.x (Unity 6)AllSupported

Platforms: Windows, macOS


Contributing

Contributions are welcome! Please see CONTRIBUTING.md for guidelines.

  • Report bugs and request features via GitHub Issues
  • Submit pull requests against the main branch
  • Add new tools by following the [BridgeTool] pattern above

Acknowledgments

Unity AI Bridge is derived from Unity-MCP by Ivan Murzak (Apache License 2.0). See THIRD_PARTY_NOTICES.md for details.

License

Apache License 2.0

来源:README.md,提交 546583b

工具

0
SourceWeft 尚未在沙箱中启动过该软件包,因此其工具未知。

版本历史

1
  1. v1.0.0最新Sep 16, 2026

评分

已满足适用检查项的 50%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分