reddit · SideProject · 10-13 06:50 69 分 ← 返回列表

I built a check that stops AI coding agents before they delete, force push or deploy

For weeks I read stories of coding agents doing things nobody asked for, told by the people it happened to. A terraform destroy that wiped 2.5 years of student work. An agent that pushed the deletion of 723 files to a main branch. A "yeah go ahead" that cost someone 29 folders of unsaved work. A rule in CLAUDE.md , written in capitals, that the agent ignored anyway. Every time, the rule existed. Nothing checked the action itself before it ran. So I built Paidwen ( https://paidwen.com ).

五维评分

付费意愿

需求真实度

个人可落地

竞争空间

新鲜度

AI 分析

痛点：AI 编程代理可能在未授权情况下执行删除、force push、部署等破坏性操作，现有规则文件无法在执行前拦截。

目标用户：使用 Claude Code、Codex 等 AI 编程代理的开发者、DevOps 工程师及需要保护生产环境和敏感仓库的小团队。

证据：A terraform destroy that wiped 2.5 years of student work. An agent that pushed the deletion of 723 files to a main branch. A 'yeah go ahead' that cost someone 29 folders of unsaved work. A rule in CLAUDE.md, written in capitals, that the agent ignored anyway. It's free to try, no card.

变现：Freemium：基础本地规则免费，团队版按席位或代理数订阅，提供 Slack/邮件审批、审计日志、策略管理、SSO、多代理支持和企业合规功能。

MVP：2 周内做出 Claude Code/Codex hook 集成，支持用自然语言或 YAML 定义 allow/block/ask 规则，在工具调用前拦截危险命令，并通过 Slack/邮件发送待批准操作与审计日志。

竞品：Claude Code hooks/permissions、OpenAI Codex CLI approval modes、Cursor command allowlist/denylist、Git hooks/CI branch protection、HumanLayer

摘要：发帖人因看到多起 AI 编程代理破坏性操作案例，例如 terraform destroy 抹掉 2.5 年学生作品、向 main 分支推送删除 723 个文件，开发了 Paidwen。它接入 Claude Code/Codex hook，在执行前检查工具调用，并可通过 Slack/邮件人工批准。当前免费试用、无需信用卡，原帖未给出付费意愿或收入数据。