科技纪年/Tech Craft Chronicles

深夜十一点,你的Mac突然多了两个不吃饭、不摸鱼、不要五险一金的AI打工人:一个埋头写代码,一个专挑它茬儿,你只用喝着奶茶当甩手掌柜。合上电脑盖子它们照样干活,重新打开一切原地不动——像极了你梦寐以求却从没遇到过的靠谱同事。唯一的缺点是:它们偶尔也会一本正经地把bug写得很有感情,需要你时不时探个头,假装很忙地瞄一眼。 At 11pm your Mac quietly gains two coworkers who never eat, never slack off, and never ask for benefits: one writes code, the other nitpicks it, while you sip tea and pretend to supervise. Close the lid and they keep grinding; reopen it and everything's exactly where you left it — the reliable colleague you've always wanted but never had. The only catch: they occasionally write bugs with real conviction, so you still have to peek over now and then, looking busy. 实操案例: 用Herdr同时开两个完全免费、跑在本地的AI Agent,一个写Python天气脚本、一个负责审查,合力产出一张天气趋势图。 / Using Herdr to run two completely free, fully local AI agents side by side — one writes a Python weather script, the other reviews it — together producing a temperature chart. 1. 先请此神器落地 Bring Herdr to life Mac用户一行命令装好,或者用Homebrew也行。 / Install with one command, or use Homebrew. curl -fsSL ht...

  一个给AI Agent装大脑，一个给穷学生装印钞机：前者真省钱，后者真免费但有坑。/One gives your AI agent a brain; the other gives broke developers free APIs. The first saves tokens; the second is free—with strings attached. 1. Codebase-Memory-MCP：别让Agent天天翻文件/Stop Your Agent From Reading Files All Day 传统Agent经常 grep -> read file -> grep -> read ，像实习生翻档案翻到凌晨；Codebase-Memory-MCP 用纯C写的MCP Server，把整个仓库索引成持久化SQLite知识图谱。/Traditional agents loop through grep -> read file -> grep -> read ; Codebase-Memory-MCP builds a persistent SQLite knowledge graph instead. 155种语言、tree-sitter + LSP；Linux内核2800万行、7.5万文件约3分钟索引。/It supports 155 languages with tree-sitter and LSP; the Linux kernel benchmark is about 28M LOC and 75K files in roughly three minutes. 14个MCP工具，Cypher查询 <1ms ，支持调用链、死代码、架构及HTTP/gRPC关系。/It provides 14 MCP tools, sub-1ms Cypher queries, call chains, dead-code detection, architecture and HTTP/gRPC links. .codebase-memory/graph.db.zst 可提交Git共享； localhost:9749 可看3D图谱。/The compressed graph can be com...

不用买 VPS，也不用让 M2 电脑熬夜加班：让 AI 写代码，Codespaces 提供云端工作区，Tailscale 把机器安全串起来，GPU 偶尔负责重活，最后把成品丢进免费主机长期展示。唯一的小脾气是免费资源都有额度，但钱包可以继续保持安静。 Your M2 Mac can finally stop doing all the heavy lifting. AI writes the code, Codespaces provides the cloud workspace, Tailscale connects everything privately, a free GPU handles the occasional heavy job, and TinkerHost keeps the finished site online. Free clouds have limits, but at least your wallet gets a nap. 实操案例 这个例子做的是一个“免费 AI 实验室展示站”：AI 在云端帮你写一个简单网站，Codespaces 负责开发，Tailscale 提供私网连接，免费 GPU 做一次 AI 推理实验，最后把网站发布到 TinkerHost。/This example builds a “Free AI Lab” website: AI helps write it in the cloud, Codespaces provides the workspace, Tailscale provides private connectivity, a free GPU runs one AI inference experiment, and TinkerHost publishes the final site. 1. 准备免费账号/Create the free accounts 准备一个 GitHub、Tailscale、MonkeyCode 和 TinkerHost 账号，全部先使用免费方案。/Create GitHub, Tailscale, MonkeyCode, and TinkerHost accounts and start with their free plans. 2. 创建 GitHub ...

软件不会自己变聪明，但我们可以给 Agent 装上“手”：CLI-Anything 把 GUI 软件变成 Agent 能用 CLI + JSON 驱动的工具；Loop Engineering 则让系统自己决定何时跑、何时停。一个负责“手”，一个负责“脑回路”——两者叠起来，才像真正的 Agent 工程。 方案 A：CLI-Anything / Turn GUI Apps into Agent Tools 港大 HKUDS Data Intelligence Lab 开源的 CLI-Anything，可以把带源码的软件自动生成 Agent 友好的 CLI 外壳（harness），让 Claude Code / Codex 不用疯狂点鼠标，而是用命令行 + JSON 操控真实后端。 软件源码 / GitHub 仓库 → 7 阶段流水线（分析→设计→实现→测规→写测→文档→发布） → 可 pip 安装的 CLI（如 cli-anything-blender） → Agent 用结构化命令驱动真实后端（bpy、GEGL 等） 已有 Blender、GIMP、LibreOffice、OBS、Inkscape、Kdenlive 等 harness，并有 CLI-Hub 可浏览/安装。挂上 Claude Code 等插件后，一条 /cli-anything 就能开跑。 免费吗？/ Is it free? 项目开源免费（MIT），但生成 harness 要跑 Claude 等强模型，API / 订阅费用自己承担；目标软件也得先装好。 准备 / Requirements Python 3.10+；GitHub 账号；Claude Code（主路径）或 Codex / OpenCode 等；模型 API Key 或订阅；目标软件本体。一般不强制银行卡给 CLI-Anything 本身，真正花钱的是云厂商/模型商。 方案 B：Loop Engineering / Let the System Prompt the Agent Loop Engineering 不是某个 App，而是一种 2026 年流行的 AI 工程方法：别再自己一句句 prompt Agent，而是设计一个外层循环，让它发现任务、调用 Agent、验收、记状态，再决定下一步。 Loop engineering = ...

不用 GitHub、不花一分钱,用一个 Discord 要来的邀请码开通 DomCloud 免费主机,然后用官方部署脚本上线一个带 HTTPS 的静态个人主页,并把 SSH、数据库和网页 VS Code 全部验证一遍。/ With no GitHub and no money, use an invite code from Discord to activate a free DomCloud host, then deploy a static personal homepage with HTTPS using the official deploy script, and verify SSH, the database, and browser-based VS Code along the way. 完整案例: macOS M2 16G 在这里只负责浏览器和一个终端窗口，网站跑在对方机房。Your M2 with 16GB only supplies a browser and one terminal window; the site runs in their datacentre. 1. 注册账号 / Register an account 打开 my.domcloud.co/login，用 Google 登录最省事。Open my dot domcloud dot co slash login and sign in with Google, which is the least hassle. 用第三方登录可以跳过邮箱验证；用邮箱注册则必须先点验证链接才能建站。Third-party sign-in skips email verification, whereas registering by email means you must click the verification link before creating any website. 2. 先看清免费版给你什么 / See exactly what the free plan gives you 1.5GB 存储、2GB 月出站流量、SSH、Git 集成、数据库，以及和付费版一样的功能面。1.5GB storage, 2GB monthly outbound traffic, S...