Community posts

Claude Code now has mods: TypeScript functions that run inside the agent and can rewrite prompts, guard tool calls and redraw the interface. They also run with your permissions and no sandbox. Here's what they do, and what to check first. https://x.com/i/article/2106421890797428737

@robfreshwater @ClaudeDevs Highlighting the moments Claude is actually waiting on you is such a useful mod, those questions get buried mid-run. Having Claude make the demo video too is a nice touch

Claude Code mein har baar /usage type karna padta tha. So I asked Claude to write a mod for it. Now my usage sits right above the prompt, and when a session starts or the limit runs out, it plays KBC sounds 🔊 @claudeai @ClaudeDevs @bcherny @shantanugoel

Built my first mod with @ClaudeDevs: it blocks me when my English is sloppy 😁 Claude understands me through any typo, so I stopped caring. The bad habits followed me into human conversation Now Claude won't reply until I write properly. Time to show AI some respect 👮

Having fun with the new mods in @ClaudeCode 🛠️ I built agent-swarm because the default subagent view in Claude Code doesn't show you much: no model or effort, no cost, no idea which tool each agent is running. Just Agent(…), spinner, and tokens. Now every subagent gets a live row above the prompt, and you can click any one to follow its whole run. https://github.com/jeffarese/agent-swarm @claudeai @ClaudeDevs

Repo (MIT, projeto independente): https://github.com/Jakeschincariol/arena-skill A skill é de graça, mas gasta uso do Claude Code. Modo cheio = 595 calls de sub-agent; --quick = 91 (números da própria skill, não é benchmark). Como plugin o comando vira /arena-skill:arena. Copia a pasta se quiser plain /arena.

@bcherny I’d be interested in accessibility mods: fewer moving elements, larger status text, and keyboard shortcuts for the actions someone uses most. Small changes like those could make an agent much easier to work with for hours.

夏前にちょっとしたツールを作って満足してたけどClaudeが凄いらしいから来月から課金してみよう ただ、今特に作りたいものがない もう少ししたら好きなゲームが出るからそのMOD製作とかならチャレンジしてみたいんだけどな…

Anthropic just let users rewrite Claude Code itself. They call it mods: small TypeScript functions that change a prompt, add new UI, or replace a built-in feature. The catch: mods run unsandboxed, with full access to your machine. The tool is now yours. So is the risk.

e o novo recurso do Claude Code Mods me permitiu melhorar a integração entre Codex e Claude com uma interface personalizada dentro do chat

@lydiahallie Yeah! Built a mod that lets my OpenAI dot coordinate tasks with Claude Code agents: send work, wake idle sessions, and read replies. Beyond dots, other agents and local automations can hand tasks to Claude Code and collect the results. https://github.com/ParsingMAINFRAME/Dots-Claude-Code-Bridge

Claude Code Mods just landed (≥2.1.287; pin 2.1.288). JS/TS plugins that run inside Claude Code: rewrite prompts, guard/retry tool calls, draw panes. Deeper than skills or MCP alone. https://claudekit.io/en/updates/claude-code-mods/

My first mod this morning to test Claude Code mods. Made a tiny pet that naps above the prompt while Claude is idle and wakes up when Claude starts working. Cat hunts flies, dog gets the zoomies, and the otter never drops its rock. Useful ones will come, this was just a test

Claude Code just became a platform. Anthropic shipped mods: small TypeScript functions that change behavior and UI, installed with /plugin. The announcement did 2.63M views in 46 hours. The extension-marketplace playbook is repeating. The moat is moving to the ecosystem, not the model. Ship with an extension story on day one.

@lydiahallie middleware is exactly how I'd explain it. first mod everyone should write: log every tool call with a timestamp. boring, but you finally see what the agent actually did

蹲到个新东西：Nous Research 的开源 Hermes Agent。今天在时间线上刷到好几条，持久记忆加闭环自我学习这个组合有点意思，社区那边说昨天一天就合并了 175 个 PR，几十个新插件进了目录。开源 agent 这条赛道是真热闹，Claude Code 刚出 Mods 系统，这边 Hermes 也在疯狂迭代，找时间实际跑一下再说

这个对比写得太细了。最有意思的结论：Claude Code 在补 Pi 的课，Pi 在学 CC 的克制。Mods 把插件开发做成自然语言+热重载的默认闭环，门槛是真降下来了——插件生态要从极客玩具变成大众能力了

Claude Code 在 2.1.287 版本放出了 Mods 系统，我第一时间把官方发布说明和 Addy Osmani 写的开发教程从头到尾扒了一遍 给我的第一感觉是，Claude Code 终于开始长成 Pi 的样子了 😂 Pi 也越来越像 CC 了 hhhhh 之前我从 Claude Code 转去主力折腾 Pi，很大一个原因就是早期的 Claude Code 只开放了「两头」：开头让你写写 CLAUDE.md 和 Skill，末尾让你接几个 MCP，或者在 settings.json 里配个 shell hook，每次事件触发还得起子进程通过 stdin/stdout 传 JSON 至于中间的执行过程完全是个黑盒，既没法在会话里留状态，也改不了界面，更别提替换内置功能 结果这次 Anthropic 直接用 Mods 把中间这条链路全打开了，甚至在官方博客里明说：现在连内置的 /diff 面板和 AGENTS.md 支持都已经改用 Mods 实现，放在公开仓库的 mods/ 目录下，后续还会把更多内置功能迁出来，让大家可以把 Claude Code 削成一个最小内核，只按需装回自己想要的能力 「薄内核 + 全插件 Harness」这条路，现在连官方大厂也彻底认了 不过把 Claude Code Mods 的 API 和我平时天天写的 Pi Extension 放在一起对比，两边的底层设计思路其实有几个很有意思的差异 👇 1️⃣ 拦截方式：一个是洋葱圈套娃，一个是流水线卡口 以前在 Pi 里理解插件，脑子里是一条从「用户输入 -> 单步思考 -> 跑工具 -> 循环结束」的单向流水线 Pi 把执行前后的关节拆成了各自独立的卡口，比如刚输入时、发给模型组装上下文前、工具执行前、工具执行后，你想改哪一步，就直接去守哪个路口 而 Claude Code Mods 走的是「洋葱圈套娃」模型 所有已安装的 Mod 按加载顺序套成一圈又一圈，每次有事件发生（比如准备调命令行、准备画界面），事件就从最外层的 Mod 一层层往里传，到底部交给 Claude Code 执行完，结果再一层层往外退 在任意一层里，你的 Mod 都可以选三种动作： 先放行再观察：让事件先往下走，等内核跑完拿到结果了，再顺手记一笔日志或算个用量 半路改写：把传进去的危险命令改安全点再放行，或者等工具跑完后，把输出里的敏感密钥抹掉再交还给模型 直接接管：根本不往下传，当场替内核拒掉这次工具调用，或者干脆用你自己的逻辑替换掉官方的内置命令 这种套娃顺序还被 Anthropic 顺手拿来做了一层企业安全防御：团队版机器默认把官方的内置安全 Mod 排在最外层第一个加载，后面员工自己装的插件就算想篡改权限规则，也会在最外层被直接按住 2️⃣ 运行边界：受控通道与 10 秒限时 vs 本地直通 在 Pi 里写插件，代码直接跑在本地环境里，调系统能力和弹窗都非常直接 Claude Code Mods 则把插件放进了一个不直接暴露底层全局变量的受控环境里，读写文件、跑外部命令、画界面，全部得走官方给的统一宿主通道 但别误会，这只是接口形式上的收敛，并不是真正的系统级安全沙箱 官方专门提醒，Mod 依然拥有跟 Claude Code 本身完全相同的电脑读写与命令执行权限，乱装来路不明的插件照样危险 这里有个很抠细节的设计：为了防止烂插件写死循环把主程序卡死，Claude Code 给每次事件处理设了 10 秒 的运行上限，但调用官方通道等待外部响应的时间不计入这 10 秒里 所以在官方给的高危命令拦截示例里，当拦住删库命令需要挂起等待人类按键确认时，插件里不得不用每隔 0.25 秒调一次系统休眠的办法来「合法续命」等用户拍板；而在 Pi 里，直接调一下确认弹窗就能挂起等待，不用绕这个弯 3️⃣ 开发体验与上下文控制：嘴遁热重载 vs 历史与投影分离 这次 Claude Code Mods 体验上最讨巧的一点，是把「自然语言写插件 + 会话内热重载」做成了默认闭环 你甚至不用离开当前终端去翻文档建文件夹，直接在对话里跟 Claude 描述你想在输入框上方看什么指标，允许热重载后，当前回合一结束，新写的插件就直接画在屏幕上了，后续每改一句话都会原地刷新 为了防止每次热重载把插件刚才攒下的统计数据洗掉，Claude Code 让宿主统一替插件保管状态，还配了强制的类型检查：只要界面在渲染时读过某项状态，后面数据一更新，界面就会自动重绘，完全不用自己操心刷新时机 而 Pi 在状态和上下文这块走的是另一条硬核路线：它把保存在本地磁盘上的完整聊天流水账，跟每一轮真正发给大模型的上下文切片彻底分开了 在 Pi 里，插件可以在发送请求前把几千行过期报错日志在内存里压成一句话，模型这轮只吃浓缩后的摘要，但本地磁盘里的完整历史记录毫发无损，在精细手术式裁剪长会话上下文这件事上，目前还是 Pi 更顺手 4️⃣ 界面扩展：一套组件跨终端和桌面端 vs 极客终端 TUI 以前我们改命令行工具的界面，基本就是跟终端字符画死磕 但 Claude Code 现在要同时兼顾终端 CLI、桌面端应用甚至 IDE 扩展，所以它把「画界面」也变成了一个可拦截的事件，开放了输入框上方横幅和右侧分栏面板等位置 你在插件里声明一个文本框或按钮，在终端里就画成字符框，到了桌面端就自动渲染成原生卡片 它甚至做好了窄屏自动降级：比如你的插件想在右边开个分栏面板展示代码对比，结果当前终端窗口太窄放不下，系统就会告诉插件分栏失败，同一套界面会自动降级画到输入框正上方，不需要写两套界面代码 5️⃣ 自动化闭环：Pi 的强插队自愈 vs Claude 的四个标杆玩法 在怎么帮开发者省心这件事上，两边的发力点也很有代表性 Pi 最爽的组合拳是主动往主循环里「插队塞指令」，比如检测到单测跑挂了，插件可以不等模型交卷，当场把报错拍回模型脸上并强行开启下一轮思考，逼它继续改代码直到测试全绿才放行 而 Claude Code 这次随发布一起放出的四个官方标杆 Mod，则把「可观测性」和「防手滑」做到了极致： Token Weather（上下文天气预报）：总共就 80 行左右，在输入框正上方画一行实时天气，上下文占用低于 25% 是晴天，50% 以上变阵雨，75% 以上变暴风雨，到 90% 直接亮红字预警即将压缩，旁边还带着最近 12 轮的用量走势小柱状图和上一轮刚涨了多少 Token Blast Radius（爆炸半径预览）：只要模型想跑删目录、硬重置 Git、强推或者数据库迁移，它就先把命令按住，在后台偷偷跑一遍预检查，直接在面板上列出来「这行命令下去会删掉哪 9 个文件、共计多少 KB」，让你按数字键 1 放行或者按 2 拒掉并把原因退回给模型 Replay Theater（改动回放剧场）：在模型改文件前先把旧内容存一份，等这一轮改完几个文件后，按一下 r 或者敲 /replay，就能在侧边栏像翻幻灯片一样一步步回放刚才到底动了哪些代码 You should know（旁路监督子智能体）：官方在 2.1.287 里新塞进去的内置可选 Mod，开启后会在后台分叉出一个旁路子 Agent，专门盯着主 Claude 的输出，提醒那些主线容易顺手忽略的问题 整体看下来，如果你喜欢彻底掌控本地运行时、自己精细裁剪每一轮发给大模型的上下文投影、用强插队指令编排自愈循环，Pi 依然是目前最自由、最纯粹的极客 Harness 但如果你日常主力用 Claude Code，2.1.287 的 Mods 绝对值得立刻升上去玩玩，不用自己手搓代码，直接让 Claude 现场给你生成一个 Token Weather 和 Blast Radius 热重载挂上，平时盲开上下文和怕删库的焦虑立马就能消掉一大半 最后用一张表把两边的核心差异收个尾：⬇️

/config も、自作 Claude Mods が表示できる 真の意味で、Cladue Codeをカスタマイズできて楽しい機能 もはや、汎用的な実行環境ではなくて、個々人に併せてハーネス自体を最適化していくのが重要

anyone else prefer having the plan on the side in Claude Code CLI? created a mod so I read it and give feedback without scrolling up and down https://github.com/ellenrchen/claude-plan-pane

claude code lets you mod the UI now so naturally i used it to stare at my usage limits 24/7 also added a "what's happening?" button for when it's been cooking for 2 minutes and i start getting nervous it also shows how much followers i have on X ... useless feature i know but I’m growing 😅

YOU CAN NOW *MOD* CLAUDE CODE Here's how to build your own: First, to help you understand the scope of this thing: You can add almost anything. Multiplayer Doom, a spinner that turns into a cartoon, a samurai that fights every tool call  - all integrated directly into the tool. Okay, but what is a mod? It's just a file that runs inside Claude Code comprised of these four pieces: 1️⃣ Trigger: the moment it runs (a session starts, you send a prompt, Claude runs a command). 2️⃣ Action: watch it, change it, or block it. 3️⃣ Place: where it shows up (a band above the prompt, a side panel, a status line, a toast). 4️⃣ Data: what it can reach (your files, the shell, any web API, a call to a model). Snap those together and you can build all kinds of useful tools: → progress bars for every task, which flag when Claude is waiting on you → a button that checks every number in your report against the source spreadsheet → live signups in a panel right next to your work → a tracker for where Claude spends its time And, of course, Claude can build it for you. Just describe what you want ("make a mod that checks my numbers"), Claude writes it, and it loads while you keep working. Still not sure what to ask? I put together this free resource for you with five templates and 12 prompts to experiment with. ➡️ http://github.com/carlvellotti/mods-starter Show me what you build!

Anthropic just gave Claude Code extensions that run with full access to your machine — no sandbox. Mods can rewrite prompts, block tool calls, read your API keys. Anthropic's own docs warn: install only from trusted sources.

Made a Claude Code mod that puts one dim line under every reply: Session: 39% used · this turn: +2% So you can see which prompts eat into your 5-hour limit. On pay as you go it shows dollars instead. Works in the terminal and the desktop app.

Claudeくんは今日2つのゲームの和訳をこなし、機能しないMODを修整し、ツール作りも頑張った。 凄いねぇ。

Claude code Mods好用爱用还要多用，半个晚上就把band利用起来了。最后还跟我C调了一下情，当硅基生命统治地球时，想必他能饶我一条狗命吧😂😂😂

Built a bridge so my OpenAI dot can manage Claude Code agents: assign work, wake idle sessions, and check their replies. Requires the bridge plugin and a local mailbox. Still experimental, but working on my Mac. https://github.com/ParsingMAINFRAME/claude-mailbox-bridge My dot now manages a team Claude mod

Claude Code'a bir kedi ekledim 🐈 Claude düşünürken o da düşünüyor. Bash çalışınca terminale pati atıyor, hata olunca ürküyor. Mama vermeyi unutursan acıkıyor. Claude Code'un yeni mod sistemiyle yaptım. Kodu açık, kurulum README'de. Link'i aşağıda yanıtlarda paylaştım.

Claude Code's new plugin system can rewrite its own prompts, intercept tool calls mid flight, and reshape the interface, and none of it runs in a sandbox. Mods landed in version 2.1.287 on October 1, enabled by default, though the npm stable tag was still pointing at 2.1.285 a day later, so plenty of people don't have them yet. A mod is a JavaScript or TypeScript file that hooks into Claude Code's internal event stream: it can observe a tool call, modify it, or take it over outright. That's a deeper hook than settings files, CLAUDE.md, or MCP servers ever gave you, because it sits on the same layer that controls prompts and permissions rather than beside it. Anthropic shipped one of its own on day one. You Should Know is a built in mod that spins off a side agent to watch Claude's output and flag anything important you might have missed, installed with a single slash command. The announcement pulled 979K views and 852 reposts in a day, which tells you people have wanted this hook for a while. Here's the part that got less attention than the feature list. Mods run with full user permissions. No sandbox. They can read and write any file, launch processes outside Bash's isolation, make arbitrary network calls, and touch your environment variables, meaning your API keys. Anthropic's own guidance is blunt about it: install mods only from authors and marketplaces you trust. And because Claude Code can write and execute a mod mid session on its own, the audit trail isn't only what a developer shipped, it's what the model decided to author and run while nobody was reading it. That's a category of security review nobody's process accounts for yet: auditing code your own AI wrote and ran against itself, in production, before a human read a line of it. Enabled by default, with no sandbox, on day one. The ecosystem is one supply chain incident away from finding out how big a problem that is.

Claude Codeに、自分専用の画面や危ないコマンドの前の確認を足せる「mods」が加わりました。頼めばClaudeが作れる一方、隔離されずにAPIキーにも届くため、入れる前の確かめ方と止め方まで整理しています。 https://x.com/aicreatorpath/status/2106416238155145681