[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fkg0y5h8dknx8":3},{"code":4,"message":5,"data":6,"timestamp":74},10000,"success",{"full_name":7,"position":8,"description":9,"language":10,"stars":11,"forks":12,"stars_gained":13,"age_days":14,"hn_threads":15,"one_liner":46,"summary_md":47,"topic":48,"github_topics":51,"license":70,"homepage":71,"updated_at":72,"history":73},"trycua\u002Fcua",5,"Scale computer-use 2.0 with open-source drivers, cross-OS sandboxes, and benchmarks for training, evaluation, and data generation.","Rust",30861,2251,4875,631,[16,22,28,34,40],{"id":17,"title":18,"points":19,"comments":20,"posted_at":21},49259339,"Apple Silicon and macOS VMs: Faster LLM Inference with llama.cpp",307,41,"2026-08-11",{"id":23,"title":24,"points":25,"comments":26,"posted_at":27},47936312,"Show HN: Drive any macOS app in the background without stealing the cursor",192,43,"2026-04-28",{"id":29,"title":30,"points":31,"comments":32,"posted_at":33},43773563,"Launch HN: Cua (YC X25) – Open-Source Docker Container for Computer-Use Agents",172,73,"2025-04-23",{"id":35,"title":36,"points":37,"comments":38,"posted_at":39},43985624,"Show HN: Lumier – Run macOS VMs in a Docker",159,52,"2025-05-14",{"id":41,"title":42,"points":43,"comments":44,"posted_at":45},49767564,"Show HN: CUA-S1 – A System One Model for Computer Use",95,10,"2026-09-19","为 computer-use agent 提供可自行掌控的 macOS\u002FWindows\u002FLinux 真实桌面与沙箱环境。","## 解决什么问题\n\n做 computer-use agent 时,真正难的不是模型,而是它脚下那套基础设施:一台能跑的机器、能打开应用并读取屏幕、能注入输入、跑完还能重置。多数团队最后拼出一堆 VM 脚本和截图循环,只能在自己笔记本上跑,也没法并行评测。Cua 把这一层拆成几个组件:托管 agent \"Space\" 的桌面应用、操作原生应用的 driver、Apple Silicon 上的 VM 管理器、沙箱 SDK\u002FCLI,外加一个小型决策模型和一个评测工具。\n\n## 有什么不同\n\n最关键的一点是 SDK 里不内置 agent。沙箱就是普通运行时,由你自己的 harness 驱动,规划逻辑完全留在你手里,不用为了用沙箱去接受一个框架。其次 Cua Driver 覆盖 macOS、Windows、Linux 三端的原生应用,在平台允许时可以后台投递输入、不抢用户指针,这在 GUI 自动化库里并不常见。第三,Lume 基于 Apple 的 Virtualization.Framework 在本机建 macOS 虚拟机——大多数 agent 沙箱方案根本给不了 macOS,而这里它是第一等目标。相比 Playwright 这类只做浏览器的方案,它更底层、平台更原生;相比通用容器沙箱,它又多出了 macOS 和桌面应用这一块。\n\n## 适合谁\n\n适合正在构建或评测 computer-use agent、需要 macOS 环境、并行沙箱或可复现桌面会话的团队,前提是你自己已有 agent 逻辑。做轨迹数据的研究者和搭建内部沙箱集群的平台工程师都能受益。如果你只需要 Linux 上的浏览器自动化,Playwright 更轻;如果你想要开箱即用的托管 agent 产品、不想碰基础设施,这里也不合适。README 很长,部分段落营销味较重,多机集群和 Windows Space 的细节明显比 macOS 单薄。\n\n## 上手成本\n\nSpaces 要求宿主机为 macOS 26 或更高,安装包已签名,但要手动授予权限。Lume 虚拟机仅限 Apple Silicon。CLI 和 SDK 覆盖 Python、TypeScript、Rust,Swift 与 Kotlin 绑定明确尚未发布。单机跑通只要几条命令,但做成真正的集群意味着准备机器、relay 和网络配置。许可证是拆开的:仓库本体 MIT,而 Cua Spaces 应用是 FSL-1.1-MIT 的 source-available,两者并不等价。Spaces 对个人免费,Pro 和 Teams 收费,商业使用前需要做许可证评估。\n\n## 点评\n\n这是一个 2025 年初才创建的新项目,已经积累约 3.08 万 star、2248 fork,最近一个月净增约 4875 star,在 Rust 月榜排到第 5。对于一个组件这么多的技术栈,这个速度相当快;而它始终没上过日榜,说明更像持续爬升而非单日爆发。架构方向是对的——小颗粒、可组合、不对你的 agent 做假设,而 macOS 虚拟机能力是少数竞品提供不了的。风险在成熟度和整合度:多个组件还在 1.0 之前,MIT 与 FSL 混合授权加上付费档位,意味着基于 Spaces 做产品前必须先看商业条款。如果 computer-use agent 是你路线图的核心,值得认真评测;如果桌面自动化只是附带功能,就不必了。",{"slug":49,"name":50},"ai-engineering","AI 工程",[52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69],"agent","ai-agent","apple","computer-use","computer-use-agent","containerization","cua","desktop-automation","hacktoberfest","lume","macos","manus","operator","swift","virtualization","virtualization-framework","windows","windows-sandbox","MIT","https:\u002F\u002Fcua.ai","2026-10-24T16:05:49Z",[],1792868545]