一份技术指南介绍了如何在Windows沙盒中启动Minecraft Java版，并通过运行在沙盒内的cua-driver MCP服务器，让具备视觉能力的AI Agent驱动游戏1。该指南涵盖本地与Fleet两种运行路径，要求使用cua-sandbox 0.3.3或更新版本，以支持Fleet Windows环境及镜像注册表等相关功能1。在启动时间上，本地沙盒热启动约需30秒，而默认配置为4 vCPU和4 GB内存的Fleet Windows沙盒就绪则需约3分钟1。
在图形渲染方面，由于Minecraft 1.17及更高版本需要OpenGL 3.2，而沙盒GPU仅提供OpenGL 1.1，因此必须使用Mesa3D的opengl32.dll (llvmpipe) 软件OpenGL驱动1。该环境的MCP接口共包含55个工具，其中获取桌面状态、列出应用、列出窗口、获取窗口状态、点击、双击、输入文本、按键、快捷键、启动应用、置顶、滚动和拖动等工具成功运行，而获取屏幕尺寸、获取无障碍树、获取配置、检查权限、获取光标位置和缩放等工具被拒绝1。一次从启动器到在新世界中站立的完整运行，在本地和Fleet环境下分别执行了52步和51步，耗时约20分钟1。在镜像处理上，导出基于3.40 GB覆盖层和5.62 GiB基础磁盘（64 GiB虚拟）的镜像耗时8分44秒，生成7,697,072,128字节（7.14 GiB）的数据1。构建OCI镜像耗时6分28秒（包含3分48秒导出层和1分57秒推送），层大小为7,631,366,006字节1。此外，Fleet环境签发的客户端凭证令牌有效期为900秒，并强制执行300秒的绑定截止时间1。
A newly published technical guide demonstrates how to launch Minecraft Java Edition inside a Windows Sandbox and drive the game using a visual AI agent through a cua-driver MCP server 1. The documentation outlines both local and Fleet deployment routes, covering sandbox startup, software OpenGL driver setup, MCP tool invocation cycles, image export and publishing, as well as Fleet-specific configuration and troubleshooting 1.
Because the sandbox GPU only offers OpenGL 1.1, the setup requires Mesa3D's opengl32.dll (llvmpipe) to meet the OpenGL 3.2 requirement for Minecraft 1.17 and later 1. The operation requires cua-sandbox version 0.3.3 or newer; Windows on Fleet needs 0.3.0, Image.expose() landed in 0.3.1, sb.exposed_ports in 0.3.2, and 0.3.3 brought Image.from_registry with os_type parameters and the pull-secret fix 1. A full run from the launcher to standing in a new world took roughly twenty minutes, completing in 52 steps locally and 51 steps on Fleet 1. The MCP surface comprised 55 tools; actions including get_desktop_state, list_apps, list_windows, get_window_state, click, double_click, type_text, press_key, hotkey, launch_app, bring_to_front, scroll, and drag ran successfully, while get_screen_size, get_accessibility_tree, get_config, check_permissions, get_cursor_position, and zoom were refused 1.
On Fleet, the default sandbox is 4 vCPU and 4 GB, taking about three minutes to become ready against about thirty seconds for a warm local boot 1. The exported image took 8 minutes and 44 seconds, producing 7,697,072,128 bytes (7.14 GiB) from a 3.40 GB overlay on a 5.62 GiB base disk with a 64 GiB virtual size 1. Building the OCI image took 6 minutes and 28 seconds, with 3 minutes and 48 seconds spent exporting the layer and 1 minute and 57 seconds pushing it, resulting in a layer size of 7,631,366,006 bytes 1. Furthermore, issued client-credential tokens came back with an expires_in of 900 seconds, and Fleet enforces a 300-second bind deadline 1.
评论
还没有评论，欢迎留下第一条。