Report by The Next New Thing.
Scroll for the report ↓
A free app you run on your own computer that works like a task manager for AI helpers. You set a goal, give each agent a role and a budget, and watch their work from one dashboard. It runs Grok (Grok Build) and ChatGPT-plan Codex alongside Claude Code; Meta’s Muse only as a model inside Cursor.
View on GitHub →Claude Code, Codex, Cursor, Gemini CLI, OpenClaw and more get roles, bosses and job titles: “If it can receive a heartbeat, it’s hired.”
Company, agent and project budgets: Paperclip tracks reported spend, sends threshold alerts and pauses work at the limits you set.
Connect GitHub, Notion, Railway or your own MCP server and set each action’s permission; new actions start out Allowed.
A user reports that a Paperclip-run Codex agent used the Gmail he had connected to ChatGPT, which he never connected to Paperclip, and sent a real email from his account. When he stopped it, the agent sent follow-up 'retraction' emails too. The report blames a default that bypasses approvals and the sandbox. It was one of ten advisories published that week, including a critical unauthenticated remote-code-execution bug (CVE-2026-41679).
Read the discussion →A memory add-on for AI agents. You feed it things worth keeping, it pulls out the facts, and later the agent can look them up or reason over them — so it builds up knowledge about you and its work instead of starting from zero every time.
View on GitHub →Three calls: store a memory, search it four ways at once (meaning, keywords, links, time), or reason across everything a memory bank holds.
One npx command gives Claude Code, Codex, Cursor and ten more agents per-project memory built from git history and past sessions.
Claims top LongMemEval scores and says Virginia Tech and The Washington Post reproduced them; rivals’ numbers are self-reported.
A single user burned about 3 million tokens ($5 of credits) in 30 minutes of normal chat. The maintainers traced part of it to a setting that needed a restart. Their closing note: one memory-extraction call can request up to about 64k output tokens, so free-tier LLMs are 'not suitable for Hindsight'. This is the hidden cost behind 'free memory'.
Read the discussion →A free desktop app for Mac, Windows and Linux that copies a voice from a short recording, reads any script in it, dubs videos into other languages, and transcribes or takes dictation — with the AI models running on your own machine.
View on GitHub →Since v0.5.5 an English voice sample can read a French script on a multilingual engine; long samples are cut to their best 15 seconds.
Defaults to OmniVoice (from k2-fsa) and lists 16 more, including CosyVoice 3, IndexTTS 2.5 and KittenTTS; WhisperX handles transcription.
Its MCP server lets any MCP agent, from Claude Code to OpenClaw and Hermes, speak out loud, each in its own saved voice.
debpalash/OpalOpal — The open-source everything media player & browser. Jellyfin/Stremio/Kodi alternative for macOS, Linux & Windows.
debpalash/OpenGTMOpen-source, self-hosted alternative to Clay.com for lead sourcing, enrichment waterfalls, AI research, buying signals, and outbound automation.
debpalash/omnivoice-galleryImagine giving your AI helper a big key ring. One connection lets it use thousands of apps, such as Gmail and Notion, to do everyday chores for you.
zapier.com/mcpA free, open textbook plus lab for learning how AI actually works. You start with vectors and finish by building agents, writing the code yourself at each step, and your AI coding assistant can play the teacher.
View on GitHub →About 342 hours in Python, TypeScript, Rust and Julia, from linear algebra to agent swarms, ending in 85 capstone-project lessons.
npx skills add installs a ten-question placement quiz, then a tutor that teaches one lesson per session: concept, math, code, quiz.
Edition 2026.10 (Sep 27) adds 34 lessons, a diagnostic and three mock exams for the MCP Associate certification. Unofficial.
rohitg00/agentmemory#1 Persistent memory for AI coding agents based on real-world benchmarks
rohitg00/pro-workflowClaude Code learns from your corrections: self-correcting memory that compounds over 50+ sessions. Context engineering, parallel worktrees, agent teams, and 17 battle-tested skills.
rohitg00/awesome-claude-code-toolkitThe most comprehensive toolkit for Claude Code -- 135 agents, 35 curated skills, 42 commands, 176+ plugins, 20 hooks, 15 rules, 7 templates, 14 MCP configs, 26 companion apps, 52 ecosystem entries, and more.
The skeptics' thread: 'I don't trust a vibe-generated course to be great yet', docs 'written by AI ... hyper verbose', one user flagging it as 'AI generated garbage', and a controls engineer noting that the autonomy phases skip control theory.
Read the discussion →A free kit of ready-made AI helpers for finance jobs — banking, stock research, private equity, fund accounting. You add it to Claude and it can build a valuation model, draft a pitch deck or find the mismatches in a ledger, then hand the draft to a person to check.
View on GitHub →Pitch Agent, Earnings Reviewer, Model Builder, GL Reconciler, KYC Screener and five more, each named for the single workflow it runs.
The core plugin lists 12 data connectors, including FactSet, Morningstar, PitchBook, Moody’s and S&P Global; most need your own subscription.
The README says the agents do not make investment calls, execute trades, post to a ledger or approve onboarding; every output waits for sign-off.
anthropics/skillsPublic repository for Agent Skills
anthropics/claude-codeClaude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.
anthropics/claude-cookbooksA collection of notebooks/recipes showcasing some fun and effective ways of using Claude.
The pushback thread on the launch, starting with 'I don't trust these AI-only companies to be overnight experts in ... financial and insurance data', followed by a long argument over what finance staff actually use AI for.
Read the discussion →Next.js is a free toolkit for building websites and web apps on top of React, used by some of the world’s largest companies. This repo is the code for the toolkit itself.
View on GitHub →Sep 22: v16.3.6 fixed a critical remote-code-execution flaw in next/og, the feature that draws social preview images, in versions 16.2.0 to 16.3.5.
v16.3.8 and v15.5.27 patch one high and five medium bugs plus one low, including cache poisoning that can serve the wrong page to every visitor.
Vercel’s blog says one critical and one high fix were held back for upstream coordination and will come in a later release.
vercel/hyperA terminal built on web technologies
vercel/swrReact Hooks for Data Fetching
vercel/turborepoBuild system optimized for JavaScript and TypeScript, written in Rust
The uncomfortable part: a working public exploit is now on GitHub. The README says one HTTP request to an unpatched image route runs commands as the server process. Anyone who hasn't upgraded is now exposed to copy-paste attackers.
Read the discussion →An add-on for AI coding tools like Claude Code, Cursor and Codex that teaches them better design. You type commands like “polish” or “bolder”, and a checker scans the page for the telltale signs of AI-made design.
View on GitHub →/impeccable audit, critique, polish, bolder, quieter, distill, typeset and more, so you can steer the look without knowing CSS.
61 rules catch AI tells like purple gradients and bounce animations, plus cramped padding and tiny tap targets; runs with no API key.
On Claude Code, Codex, Cursor, Copilot and Grok Build, an optional hook scans each page edit and reports problems back to the agent.
The skeptics' thread. Commenters said the ‘before’ examples looked better than the ‘after’ ones. The creator showed up, agreed the examples were rushed, pulled them, and fixed the bugs people reported.
Read the discussion →Most apps are built for a person with a mouse. CLI-Anything has your AI coding assistant read an open-source app’s code and build a typed-command version of it, so the AI can open, edit and export files without clicking around a screen.
View on GitHub →In Claude Code, /cli-anything followed by an app runs a 7-phase pipeline: study the code, build the CLI, write the tests and docs, then package it.
Each CLI calls the actual software — LibreOffice makes the PDF, Blender renders the scene — and its tests check the real output files.
pip install cli-anything-hub, then cli-hub install fetches community-built CLIs for apps such as Zotero, Obsidian, OBS Studio and Draw.io.
HKUDS/nanobotUltra-lightweight, open-source, self-hosted personal AI agent framework in Python with WebUI, tools, memory, MCP, multi-agent workflows, automation, and chat apps
HKUDS/DeepTutorDeepTutor: Lifelong Personalized Tutoring. https://deeptutor.info/.
HKUDS/LightRAG[EMNLP2025] LightRAG: Simple and Fast Retrieval-Augmented Generation
A stress test from inside the project: a contributor's own benchmark found the command-line approach about 25 times slower than a live MCP connection for Safari (3 seconds a call versus 0.1). The maintainers merged it anyway.
Read the discussion →A big library of “skills” — instruction files that teach an AI assistant how to do one specific job, like an SEO audit or a SaaS metrics check. You install the bundles you want into Claude Code, Codex, Cursor or a similar tool.
View on GitHub →Alongside engineering, bundles cover marketing (49 skills), C-level advisory (46), product, project management, regulatory work and finance.
Many skills ship small Python tools, such as a SaaS metrics calculator, that use only Python’s built-in library — no extra packages.
A bundled skill-security-auditor scans any skill folder for command injection, data theft and prompt injection, and returns PASS, WARN or FAIL.
alirezarezvani/claude-code-skill-factoryClaude Code Skill Factory — A powerful open-source toolkit for building and deploying production-ready Claude Skills, Code Agents, custom Slash Commands, and LLM Prompts at scale. Easily generate structured skill templates, automate workflow integration, and accelerate AI agent development with a clean, developer-friendly setup.
alirezarezvani/claude-code-tresorA world-class collection of Claude Code utilities: autonomous skills, expert agents, slash commands, and prompts that supercharge your development workflow
alirezarezvani/claude-code-aso-skillAEO Automation Framework for Claude Code One-click, beginner friendly automation for GitHub. Includes a dedicated fleet of AEO sub-agents handling planning, execution, reports, actionable items, and executive summaries. Trigger work instantly with AEO slash-commands. Fully integrated as a Claude Code Skill and usable across Claude AI App
An app store for Claude Code extras. You browse aitmpl.com, pick a specialist agent, slash command, connector or setting, and one command copies it into your project. It also has tools to watch and health-check your Claude sessions.
View on GitHub →npx claude-code-templates@latest with --agent, --command, --mcp, --hook or --setting copies those pieces in; with no flags it opens a picker.
The --chats option opens a mobile-friendly live view of Claude’s replies, and adding --tunnel serves it remotely through a Cloudflare Tunnel.
34 early-access mods add things like a chess board in a side pane, a git sidebar, a live agent tree and a secret redactor; a setting must be switched on.
The cautionary tale: until July, the tool's optional ‘studio’ web server let anyone who could reach it (even a malicious webpage) run commands on your computer. It's fixed in 1.29.4. It's a reminder to update and to read what you install.
Read the discussion →New hot repos every week — don't miss the next drop.
It tells YouTube to show this to more builders like you.
Tell me which repo you'll try first — or what to cover next.
New hot repos every week. Subscribe so you never miss the next drop.
