I built an open MVP for deploying AI-built software without giving agents unrestricted server access

我构建了一个开源 MVP，用于在不赋予 AI 代理无限制服务器访问权限的情况下部署 AI 生成的软件

AI can build software incredibly fast. Putting that software into real infrastructure is a different problem. An AI agent can write a working app in minutes, but there is still no good default answer for “how does this app actually get onto a server?” Either a human does everything by hand, or the agent gets unrestricted shell access to production. The first doesn’t scale. The second is how small mistakes become big incidents. AI 构建软件的速度快得惊人，但将这些软件部署到实际的基础设施中却是另一个难题。AI 代理可以在几分钟内写出一个可运行的应用程序，但对于“这个应用到底如何部署到服务器上？”这个问题，目前还没有一个好的默认方案。要么由人工手动完成所有操作，要么赋予代理对生产环境的无限制 Shell 访问权限。前者无法扩展，而后者往往会导致小错误演变成重大事故。

Guardian Cloud is an open MVP for deploying small AI-built applications without giving an AI agent unrestricted authority over the server. The positioning is one line: Deploy AI-built software with controlled execution. Guardian Cloud 是一个开源的 MVP，旨在部署小型 AI 生成的应用程序，同时无需赋予 AI 代理对服务器的无限制权限。其定位可以用一句话概括：通过受控执行来部署 AI 生成的软件。

The flow is intentionally simple: Paste a public GitHub repository. Guardian analyzes the application (Node.js apps are detected automatically; you provide the port when necessary). You see the deployment plan before any infrastructure changes. You explicitly approve the deployment. Guardian performs the controlled infrastructure actions, with bounded deployment monitoring, automatic domain provisioning and automatic HTTPS. The application is deployed, validated and exposed through its public application URL. When success cannot be proven, the system says so: a failed build results in an honest FAILED state, never a fake success. Fail-closed is a feature. 其流程被刻意简化：粘贴一个公开的 GitHub 仓库地址。Guardian 会分析该应用程序（Node.js 应用会自动检测；必要时由你提供端口）。在任何基础设施变更发生前，你会看到部署计划，并需明确批准部署。随后，Guardian 执行受控的基础设施操作，包括有限的部署监控、自动域名配置和自动 HTTPS。应用程序部署完成后，系统会进行验证并通过公共 URL 发布。当无法证明部署成功时，系统会如实反馈：构建失败会显示为诚实的“FAILED”状态，绝不会伪造成功。这种“故障关闭”（Fail-closed）机制是一项核心功能。

What the current MVP supports

当前 MVP 支持的功能

• Public GitHub repositories (公开 GitHub 仓库)
• Node.js applications (Node.js 应用程序)
• Application detection (自动应用检测)
• User-provided port when necessary (必要时支持用户提供端口)
• Deployment plan before execution (执行前提供部署计划)
• Explicit approval (明确的部署批准)
• Controlled deployment (受控部署)
• Bounded deployment monitoring (有限的部署监控)
• Automatic domain provisioning (自动域名配置)
• Automatic HTTPS (自动 HTTPS)
• HTTP validation (HTTP 验证)
• Public application URL (公共应用 URL)
• Honest FAILED state when a build fails (构建失败时显示诚实的 FAILED 状态)

Honest limitations

诚实的局限性

This is intentionally an early MVP: Public repositories only, no private repos; Node.js only for now, no attempt to support every runtime yet; No dashboard, no billing, no teams, no custom domains; No claims of universal support, zero-risk execution or production-grade universal cloud. 这刻意保持为一个早期 MVP 版本：仅支持公开仓库，不支持私有仓库；目前仅支持 Node.js，暂未尝试支持所有运行时；没有仪表盘、计费系统、团队管理或自定义域名功能；不宣称支持通用场景、零风险执行或生产级通用云服务。

Kept-honest test evidence

诚实的测试证据

Here is what the last validation round actually showed: an external Node.js sample repository was deployed and later served GET / = 200 publicly; another external Node.js repository (openshift/nodejs-ex) reached the deployment stage and failed honestly at the build step; the final end-to-end LIVE test used a small public Node.js test repository (AndersonVitaease/guardian-cloud-e2e) and reached a public HTTPS URL with GET / = 200. Claims don’t go beyond that. 最近一轮验证的结果如下：一个外部 Node.js 示例仓库被成功部署，并公开提供了 GET / = 200 的响应；另一个外部 Node.js 仓库 (openshift/nodejs-ex) 进入了部署阶段，并在构建步骤中诚实地报错失败；最后的端到端实时测试使用了一个小型公开 Node.js 测试仓库 (AndersonVitaease/guardian-cloud-e2e)，并成功通过公共 HTTPS URL 实现了 GET / = 200。我们的声明仅限于此。

Try it

试用

Guardian Cloud is open now: https://guardian-cloud-mvp.2-25-96-245.nip.io We are opening it now because we want real software and real users to expose the gaps instead of trying to predict every requirement in advance. Give AI agents capabilities. Not unrestricted authority. Guardian Cloud 现已开放：https://guardian-cloud-mvp.2-25-96-245.nip.io 我们现在开放它，是因为我们希望通过真实的软件和真实的用户来发现不足，而不是试图预先预测每一个需求。赋予 AI 代理能力，而不是无限制的权限。