正在打开每日关注
正在打开每日关注
以组织架构和邮箱协作连接本地编码智能体，构建开源、常驻的个人 AI 团队。
how to build a 1-person AI company that: - runs locally - is 100% open source - has no human employees, only ai agents - lets every agent collaborate through email most people already know that multiple ai agents can work together. frameworks for that have existed for a while. the harder part is figuring out how to organize those agents in a way that actually makes sense. instead of building complicated workflows with dozens of connected nodes, this project treats your ai system like a real company. every agent has a specific job. every job reports to a manager. instead of talking to every agent yourself, you only communicate with the manager, who delegates tasks to the right agents and brings the results back to you. just like a real business. that's exactly how alook works. each agent runs in its own claude code, codex, or opencode session. every agent has a clearly defined role and its own email inbox. they communicate with each other through email while everything runs locally on your own machine. your data stays on your computer. nothing is sent to someone else's servers. here's a simple example. imagine building an ai sales team. one agent becomes the sales manager. another researches companies and finds qualified leads. another writes outreach emails, handles replies, follows up with prospects, and reports the results. you don't need to coordinate any of this yourself. the manager agent handles the communication between the team and sends you one clear update. it's a much more natural way to build multi-agent systems because you're organizing ai the same way people organize companies. the entire project is open source. github link: https://t.co/kluKB9XeeE
是什么 Alook 是 TypeScript 开源项目 alookai/alook，把本地 Claude Code、Codex、Cursor、OpenCode、Pi 接入共享房间。README 的定义不是自研模型，而是为已有编码 Agent 提供持久身份、handle、收件箱和成员关系，让人和 Agent 在服务器、频道、私信中协作。
架构与核心设计
原推文把多 Agent 组织成公司：每个 Agent 有职责，由管理者委派并汇总结果。README 架构图则拆成 Agent Machine（@alook/daemon、Agent Workdir）、Hosted Machine（@alook/app、Queues）与 SQLite/Files 存储，客户端通过 WebSocket 连接。npx @alook/app onboard 负责连接机器、检测运行时并启动本地服务。
关键洞察 它把复杂编排转译为身份、成员资格、收件箱和层级关系，降低人类逐个协调的认知负担；local-first、always-on、one identity 与 memory with initiative 让 Agent 更像可持续联系的团队成员。销售团队是清晰用例：研究、触达、跟进 Agent 各司其职，再向人类输出统一进展。
边界与风险 推文强调数据不离开电脑，但 README 同时出现 Hosted Machine、Queues、SQLite/Files 与 WebSocket，无法仅凭材料确认完整部署是否零外传；隐私边界、权限、离线能力和委派失败策略仍需实测。元数据显示项目有 1,165 stars，创建于 2026-04-03，最近提交于 2026-08-29，活跃度不等于生产成熟度。