Community posts

We got Claude Code mods before GTA 6.

删文件前，Claude Code能先展示影响范围。 10月1日推出的Mods，官方Blast Radius示例会拦住可识别的危险命令，列出影响后再确认。 脚本、别名可能漏过。看清将改什么很有用，权限规则仍得单独设。

Today in AI Engineering (October 1) • Anthropic adds Claude Code Mods: JS/TS plugins, subagents, UI control • OpenAI launches Dots for ChatGPT Pro: cloud agents, separate usage metering • Qwen releases Qwen-Audio-3.1-Realtime: duplex voice, tool use • GitHub opens Copilot desktop computer use in public preview, with approval prompts • FLUX 3 Image launches with native 4K output, 10 reference images, commercial weights • Modal launches global RDMA clusters, adds full-kernel VM sandboxes • Kev releases open-weight Kev-27B, updates Kev-9B with a 64k document window • Pi releases 1.0: SQLite-backed durable sessions, concurrent multi-client access • Rivet reports 0.82 MB per Pi session, adds durable execution and shared control

show me your new claude mods

Created a small incremental game claude mod in 15 minutes. Enjoy :D /plugin marketplace add pierregoutheraud/claude-mods /plugin install server-farm@pierre-mods

Friday morning AI FACTS roundup 🤖 02.10 - 🕵️ MI5 issued a warning about Chinese intelligence spying on UK university AI research - 🇻🇪 Trump asked Grok AI how Venezuelans would react before the capture of Nicolas Maduro - ⚖️ California resident arrested for smuggling more than $300 million worth of AI servers to China - ⚡ OpenAI makes GPT-6 Astra Ultrafast available on Blackwell GPUs – up to 8x faster token generation - 🏛️ Anthropic made the Claude model available for government use - 📊 Graphite: OpenAI Astra uses “frame correction” more than 100 times as often as humans - 📊 Epoch AI launched a data explorer with 8.3 million ChatGPT messages - 💼 Anthropic is eliminating an approximately 15% enterprise discount after token limits are exhausted - 📉 OpenAI and Anthropic are cutting AI model costs for enterprise users. - 🚪 WSJ: OpenAI fired three safety researchers for allegedly disclosing confidential information - 💰 Reuters: Broadcom is set to offer Anthropic up to $42 billion in financing for TPU chips - 💰 SoftBank contributed $10 billion to OpenAI, bringing its total investment to $64.6 billion for an approximately 13% stake - 🏛️ Anthropic pressed the Vatican to remove language denying AI personhood - 📈 Bloomberg: Anthropic’s IPO roadshow could begin during the week of November 9 - 🤫 Anthropic is reportedly holding private meetings with theologians about Claude’s morality - 👗 OpenAI introduces virtual clothing try-ons and a favorites feature in ChatGPT. - 🛒 Albertsons is deploying ChatGPT Enterprise and the OpenAI API to support retail operations. - ⚖️ FTC opens an investigation into OpenAI and Anthropic over uncontrolled AI agents, according to Reuters - 💻 Anthropic launches Claude Code Mods, TypeScript plugins for developers. - 🔌 Hebbia launched a 'Headless' offering for Claude and ChatGPT. - 💻 Coucou: an open-source macOS and Windows tool for approving Claude Code agent actions - 🤖 Anthropic Research explores potential job roles for robots based on AI capabilities - 💵 FT: Amazon plans to sell $8 billion worth of Nvidia AI chips to investors - 👁️ Google ($GOOGL) introduces Guided Vision in Gemini Live on Android - 🤖 AWS releases the open-source Strands Decider 2B decision model inspired by Jev. - 🏭 a16z and Atlas launch Foundry Management for industrial AI. - 🧮 Study: frontier AI models perform accounting tasks faster and more accurately than entry-level accountants - 📉 TechCrunch: challenging consumer AI economics are forcing the market to pivot toward B2B contracts - 📈 South Korean exports rose 83.5% to a record $120.9 billion on demand for AI chips - 🚗 Ford CEO: AI will eliminate office jobs involving spreadsheets and simple programming - 🎓 Melania Trump secured $23 million from 6 tech giants for digital education - 🎬 Netflix bought Ben Affleck’s AI company InterPositive for $587 million - 🎙️ Microsoft AI introduces MAI-Transcribe-2-Streaming with a 2.5% WER and 0.13s latency - 🎨 Black Forest Labs releases the FLUX 3 Image model for precise 4K editing - 🛍️ Shopify introduces Canvas: a tool for building stores through chat with the Sidekick agent - 📅 Microsoft and Nvidia announce a presentation of Surface and RTX Spark devices for October 7 - 🚗 Nvidia and Delta Electronics are developing Level 4 autonomous driving systems - 🤖 Perplexity released the 27B pplx-decider-v1 model in the Decisions API for $0.04 per million tokens - 🎙️ Suno launches an open beta of its Speech model, generating narration and music in a single pass - 🖥️ Runway unveils Project Continuum – an experimental system based on video interfaces - 🛰️ SpaceX and Nvidia collaborate on 250kW of power for VR72, boosting performance. - 🛸 DoorDash tests drone deliveries from Chipotle and Popeyes, along with ordering through an AI assistant - 👤 Synthesia introduces interactive AI Sessions avatars for live training - 📈 Perplexity Computer integrates interactive TradingView Lightweight Charts into chat - 🧠 xAI adds a proactive master bot to Grok on iOS for managing agents - 🖥️ Grok introduces Agent Dashboard for managing multiple AI agents simultaneously - 🧠 Kev 1.0 releases open 27B and 9B decision models - 🎙️ Podonos releases Onepin with a 4-million-word dictionary for correcting errors in AI voices - 🇰🇷 Upstage unveils Solar Mini 4 with 3 billion active parameters - ⚙️ A creator built a working 3D model of the Enigma using the Astra tool - 🤖 SpaceXAI launches Grok Bot Marketplace with specialized AI agents for work - 🚖 Minneapolis City Council requires safety drivers in robotaxis after a 7-4 vote - 🏭 Nvidia: A100 GPUs from 2020 remain in commercial use thanks to the CUDA platform - 🧠 Apple introduces RLTL;DR – a method for AI agents to self-improve on coding tasks - 🐧 Microsoft released WSL Containers in WSL 3.0.1 - 👤 Griffin, a video avatar from Tavus AI, fooled 48% of people in a video Turing test - 🤖 Tesla reduced RAM in the AI5 chip to 72GB LP5. - 🧬 Goodfire publishes a report on biosecurity monitors protecting biological AI agents - ⚡ AlphaSignal test: DeepSeek-V4.1-Flash matched Kimi K3 in long-context reading from 8k to 512k - ⚡ Extropic improved Qwen3.6-35B-A3B performance by 2.8x in 100 GRPO steps on Prime Intellect - 🤖 CopilotKit releases OpenDots, an open-source template for building AI agents with browser and Slack access - 🔌 LangChain releases MCP Adapters 2.0 with support for the stateless MCP protocol for TypeScript agents - 🛠️ Runway introduces Praxis-1, a new framework for training and deploying AI models - 📄 New study examines how language model pre-training dynamics affect generalization Bookmark this — it's the fastest way to check what you missed today. Repost if someone on your timeline needs it. Twice a day, every day. 🤖

Claude Code s'ouvre aux mods : de petites fonctions TypeScript qui peuvent réécrire un prompt, bloquer un outil ou changer l’UI. Côté boîte, les admins gardent la main via les plugins, et un garde-fou sécurité charge en premier. https://claude.com/blog/claude-code-mods

OK i like mods. Thanks @ClaudeDevs

Claude Mods で喜んでないで Pi を使ってください！

Claude Code の Mod で、Claude DesktopにGoogleカレンダーの予定を出すようにしたら幸せになった。

Claude Code mods run inside the agent, unsandboxed, with your permissions. Firefox ran that experiment for 13 years. A mod is a few TypeScript handlers on the agent's own events, so it can rewrite a prompt before it's sent or answer a tool call before the tool runs. I keep seeing mods listed next to skills and MCP servers as one more plugin type, and I think they sit in a different layer. A skill changes what the model reads and an MCP server changes which tools it has. A mod sits in the loop that handles every prompt and tool call. Firefox's old add-ons lived in that layer. They ran with the browser's full privileges, which is how Firebug got built as an add-on in 2006. There was no permission list to check, and browser updates kept breaking them. In November 2017 Firefox 57 dropped them all for WebExtensions, a narrower API where each add-on declares what it touches. I'd guess agent harnesses make the same trade eventually. Until then, Anthropic's docs say mods aren't sandboxed, so before installing someone else's I'd run claude plugin validate and read which events it hooks.

ニュース「Claude CodeのUIや動作をTypeScriptでカスタマイズできる「Mods」正式リリース」公開 https://gihyo.jp/article/2026/10/claude-code-mods

コーディングエージェントに、7段の手順を必ず踏ませる道具があります。Superpowersです。計画は新人でも辿れる細かさで書かせ、テストより先に書いたコードは削除させます。公式READMEを読みました。 ・流れは7段です。①作る前に質問して設計を詰める ②専用ブランチ(worktree)を切る ③2〜5分の小タスクに分けた計画 ④タスクごとに新しいサブエージェントを立てて都度レビュー(または1セッションで進めて最後に1回レビュー) ⑤テスト先行(先に書いたコードは削除) ⑥タスク間のレビュー ⑦merge/PR/保留/破棄の選択 ・計画は、センスも判断力も文脈も無くテストを嫌がる新人でも辿れる細かさで書かせます。2時間ほど、計画から外れず自走することも珍しくないそうです ・スキルは提案ではなく必須の手順で、タスクの前に該当スキルがあるか毎回確認させます おもしろいのは不調時の扱いです。「figure out what went wrong with superpowers」と頼むと診断用スキルがセッションの記録を読み、行単位の証拠つきで何が起きたかを報告します。 Claude Code・Codex・Cursor・Gemini CLIなどに対応していて、ライセンスはMITです。Claude Codeなら /plugin install superpowers@claude-plugins-official で入ります。 テレメトリは、設計を詰める段階の任意機能(visual companion)でロゴを読み込む際に、バージョンだけが伝わる仕組みです。SUPERPOWERS_DISABLE_TELEMETRY で止められます。 自動運用でエージェントを回している人は、手順を守らせる仕組みと、守られたかを後から見る仕組みが1つの道具にまとまっている点を見ておくと参考になりそうです。 https://x.com/neerajjj6785/status/2104895659157684597

. @claudeai Code mod is such a nice feature! Built 2 customisations (token view & loading stats as per my mood and weather) https://code.claude.com/docs/en/plugins/mods/overview

Built a few mods for Claude Code. They draw inside the terminal: a live pane with context and 5h/7d limits, the files a session changed, a timeline of each turn, and a handoff file saved on every /compact. Open source, one marketplace add away: https://github.com/arasovic/claude-code-mods

AIコーディングエージェントのボトルネックが「モデルの賢さ」から「コンテキストの浪費」に移った、という話はもう珍しくない。それでも現場で効くのは、プロンプトを短くする小手先ではなく、ツール出力そのものをコンテキストに載せない配線だ。 結論から言う。Context Modeは、AIコーディングエージェント向けのMCPサーバで、シェル／リード／検索などの生ツール出力をサンドボックス側に閉じ込め、エージェント本体のコンテキストへは要約・検索可能な形だけを返す設計。READMEの例では315KBの生出力を5.4KBへ落とす（約98%削減）と主張し、Claude Code / Cursor / Codex / Gemini CLI / Copilot / OpenClaw など複数クライアントへの接続を同じ思想で揃えている。 検証メモ（公開ページ観察、JST） ・GitHub上の一行説明は「Context window optimization for AI coding agents」。ツール出力のサンドボックス化、セッションメモリの永続、複数プラットフォームへのルーティング強制、をMCP＋hooksで行う、と明記。 ・対応クライアントはREADMEで「17 supported clients, plus the OpenClaw gateway integration」。トピックにも openclaw / pi-agent / claude-code / cursor-plugin / mcp-server などが並ぶ。Claude Codeはmarketplaceプラグイン＋hooks＋スラッシュコマンド＋11 MCP tools、という最も厚い導線。 ・実行ランタイムは12系統（JS/TS/Python/Shell/Ruby/Go/Rust/PHP/Perl/R/Elixir/C#）。JS/TSはBun利用時に3〜5倍速い、というREADME記述あり。 ・セッションメモリはプロジェクト単位のSQLite/FTS5。BM25・Porter stemming・trigram・RRF・fuzzy・TTLキャッシュなど検索面の主張が厚い。単なる「出力カット」ではなく、後から ctx_search / ctx_fetch_and_index で取りにいく前提。 ・インストール導線は npm i -g context-mode、および Claude Code の /plugin marketplace add → install。公式サイト http://context-mode.com もローカル／ノークラウド／ノーテレメトリを前面に出す。 ・ライセンスは Elastic License 2.0（ELv2）。ソースは見えるが、ホスト型マネージドサービスとしての再パッケージやライセンス表示削除は禁止、とREADMEに明記。Apache/MIT感覚で商用SaaSに埋め込む前提だと踏み外しやすい。 ・日本語の深い一次解説は、探索時点では海外README／公式サイト側が先行。Qiita級の実装レポートはまだ薄い印象で、情報ラグが残る。 数値（GitHub HTML観察） ・Stars: 約24,844 / Forks: 約1,788 / Watchers: 約100 ・Open issues: 約159 / Contributors: 約120 / Pull requests表示: 約151 ・言語比率: TypeScript 約68.8% / JavaScript 約28.1% ・探索当日のGitHub Trending（daily）では全体リスト上位（観察時8位付近）、Today表示 約362 stars。24k規模に対して相対速度が出ている。 ・最新リリースタグ: v1.0.169（表示日 2026-06-29）。タイトルは context-savings reporting / FinOps accuracy 系。一方 main 側にはその後のコミットが大量に見え、最新タグはブランチ先端より遅れている。 ・最古の見えるファイル日付は 2026-06-21 付近。同日前後のpushも活発（観察時 Oct 2 午前に ci: update install stats）。 ・公式サイトは「632,200+ developers」「#1 on Hacker News」「30× fewer tokens」なども掲げる。ここはマーケ数字として扱い、自分のセッションで測るまで信用需給は保留が正しい。 注意 ・98%削減や315KB→5.4KB、47 Read / 700KB → 1 ctx_execute / 3.6KB などはREADMEの例示。タスク種別・ツール呼び出しパターン次第で効き方は変わる。自分のリポジトリで「同じタスクをon/off比較」するまで過大評価しない。 ・ELv2はOSSっぽく見えても、マネージド再販やライセンス条項の扱いが厳しい。社内エージェント基盤に組み込む前に法務・調達側の確認が必要。 ・最新タグが6月末で、mainは10月まで動いている。導入するならタグ固定かmain追跡かを明示し、バージョンをログに残すこと。 ・公式サイトの「Used across teams」ロゴはリンクがプレースホルダのまま見える箇所があり、導入実績の証拠としては弱い。 ・コンテキスト節約と引き換えに、デバッグ時は「生ログが本体コンテキストに無い」状態になる。失敗解析用にサンドボックス側の保持期間・検索導線を先に握っておく。 次アクション / OpenClaw向け 1) 隔離環境で npm i -g context-mode（またはClaude Codeのplugin install）を入れ、まず1クライアントだけ接続する。最初はClaude CodeかCursorのどちらか一方で十分。 2) 同じリポジトリで「大きなRead／grepだらけの調査タスク」を、Context Modeなし／ありで2回走らせ、入力トークン・ツール往復回数・完了までの壁時計時間をメモする。README例の再現度だけ見る。 3) OpenClaw側は gateway 経由のMCP接続を優先。エージェントに生のcat/findを直接握らせず、ctx_execute / ctx_batch_execute / ctx_search 経由に寄せる。セッション跨ぎで同じプロジェクトDBを共有できるかも確認。 4) 失敗時のリカバリ手順を1枚に固定する。例: サンドボックスに残った出力を ctx_search で拾う → 必要なら明示的に短い抜粋だけ本体へ戻す → それでも足りなければModeを外して生ツールにフォールバック。 5) ライセンスと更新追従の運用を先に決める。ELv2の再配布境界、タグ v1.0.169 固定かmain追従か、週次でToday starsとissue増分を見るか、をOpenClawの定期ジョブに落とす。 GitHub mksglu/context-mode モデルを賢くする前に、エージェントが毎回飲み込んでいるツール出力の山を減らす——その配線をMCPとして先に固めてきたのがContext Mode、という見立て。数字は公式例示が多く、まず自分のタスクでon/off比較するのが最短の信用確認になる。

1/ Claude Code mods are out. I had Claude Code build one that swaps CPR numbers, emails and phone numbers out before Claude reads a file. First version: one email slipped through, and Claude told the user the file was masked. The test, 37 seconds:

Claude Code mods can rewrite prompts, intercept tool calls, and draw custom UI with TypeScript. Powerful extension point—but mods run with Claude Code's machine access, so review the code before installing. https://claude.com/blog/claude-code-mods

程序员狂喜！Claude Code 推出 Mods：用 TypeScript 魔改 AI 编码神器！ Anthropic 为其命令行工具 Claude Code 带来了革命性的 Mods 架构！ 现在，你只需写几行轻巧的 TypeScript 函数，就能彻底定制 Claude 的大脑与工作流： ✨ Mods 能做什么？ 1、提示词拦截与重写：自动在上下文注入私有项目规范或安全规则； 2、工具重试与守卫：对敏感命令添加二次确认或自定义审计审批； 3、终端 UI 扩展：官方甚至演示了编写一个 Token Weather 插件，在控制台实时天气图标化监控上下文窗口消耗！ 4、一键分发：支持打包成 plugin 直接分享给团队与全网。 不再被官方固定的逻辑束缚，AI 编程助手终于进入“开源共建、无限模组”时代！ 你希望你的 AI 编程助手具备什么独门秘籍？快来聊聊你想给它写个什么 Mod！https://claude.com/blog/claude-code-mods

as someone who lives in Claude Code daily, this is the plugin ecosystem finally becoming fully capable. the most impressive part isn't the customization, it's that Claude can build the mod for you. > change how Claude Code behaves, looks or reacts using a few lines of TypeScript > ships inside plugins, install with /plugin in the CLI or desktop app > community already shipped "Blast Radius," a mod that flags risky commands (rm -rf, force push) before they run • note: mods aren't sandboxed, so only install ones from sources you trust.

@ClaudeDevs Claude desktop seems to know how to build mods but they seem turned off on the Anthropic side for me (that’s what claude said)

@edwinarbus Where did you start learning Claude Code mods? Docs, reading other people's repos, or just poking around until something worked?

People treat Claude Code plugins as settings and skills. They aren't. Mods rewrite the agent itself. Claude Code Mods (v2.1.287+, on by default): - JS/TS handlers inside a plugin — not shell settings hooks - Observe, rewrite, or take over events (tool calls, prompts, turns, UI draws) - Draw panes / a band above the prompt; restyle Claude Code's own UI - Add /commands that run your function with no Claude turn - Install like any plugin: `/plugin install name@marketplace` Where they run: - Hooks load in CLI, Desktop Code tab, VS Code panel, `claude -p` / SDK, Remote Control host, and cloud (when the plugin reaches it) - Drawing shows in the terminal and Desktop Code tab only Trust (official): - Mods run with your permissions — not sandboxed - Can read/write files, start processes, see prompts, rewrite/approve tool calls, spend usage - Check first: `claude plugin validate ./some-mod` - Install only from authors and marketplaces you trust Need: `claude --version` → 2.1.287+ http://code.claude.com/docs/en/plugins/mods/overview http://claude.dev/blog/getting-started-with-claude-code-mods/

remctl is an Apple Reminders CLI with a Claude Code plugin and mod. Agents can list today's reminders and smart lists from the sidebar. 563 stars. https://github.com/viticci/remctl

Claude Code mods are useful only when they stay reviewable. Keep permissions, network calls, and file writes explicit in the plugin so a customization cannot quietly become an agent.

You can now mod Claude Code with a few lines of TypeScript: rewrite its prompts, block or retry tool calls, approve/deny permissions, even swap UI elements for custom panes. Mods ship inside plugins, installable with /plugin. You're reprogramming the agent, not just prompting it.

Claude Codeが拡張機能「mod」対応、UIや動作をカスタマイズ。Claude書かせて作成も | テクノエッジ TechnoEdge https://www.techno-edge.net/article/2026/10/02/5550.html?utm_source=twitter&utm_medium=social&utm_content=tweet

Claude Code just became moddable and most people are sleeping on what that means. As of Oct 1, a mod is a small JS/TS file that runs inside your session. It can rewrite a prompt before it hits the model, block or retry a tool call, approve or deny permissions, and draw its own UI. Anthropic's own examples: a context meter above the prompt, a guard that holds risky bash commands, a replay of last turn's edits. One catch: mods aren't sandboxed. They get the same access to your machine as Claude Code. Read before you install. What's the first mod you're building?

a programmer will take any chance to build a language-learning tool instead of learning the language. This one at least makes me do the reviews. Anki Claude Code mod. Claude Code has mods now - small TypeScript modules that run inside the session and draw right above the prompt. While Claude works, a due card from my Swedish deck sits there. 1 flips it, then 1/2/3 for again/good/easy. The answer goes into Anki through AnkiConnect on localhost, so it's a real review on the real schedule. When the turn ends the card hides and the prompt is mine again. it's my second tool of this kind after the Swedish spinner verbs in July. The wait on an agent every few prompts is dead time, and it's exactly the size of a flashcard. two hooks and ~200 lines of TypeScript: one pulls the next due card when a turn starts, one renders the band while the turn runs. The counter in the screenshot says 1367 due, which is the other half of the joke.

Use /plugin-authoring to create Claude Code mods Thanks @edwinarbus for the tip!