Report by The Next New Thing.
Scroll for the report ↓
An AI agent skill that treats upvotes, likes, and Polymarket odds as the ranking signal instead of SEO. The v3 engine resolves who matters first, then fans out across a dozen walled-garden platforms and lets an AI judge synthesize. It hit #1 on GitHub Trending.
View on GitHub →Reddit, Hacker News, Polymarket, and GitHub work the moment you install — no keys.
Log into x.com, install yt-dlp, or add a ScrapeCreators key to unlock X, YouTube, TikTok, and more.
Resolves the right handles, subreddits, and hashtags before a single API call fires.
A second judge scores every result for humor and virality, surfacing the cleverest one-liners.
@ChrisLaubAI — "This feels like cheating" — a real reaction to scanning Reddit and X from the last 30 days on any topic (635 likes)
Open on X →Greg Isenberg sits down with creator Matt Van Horn for a live demo (37.8k views) · Open on YouTube (25:22 → 4:47) →
Agent context windows fill up fast with verbose tool output, logs, and retrieved chunks. Headroom sits in front of the LLM and squeezes that payload by 60–95% while preserving the answer. Ships as a library, a proxy, and an MCP server.
View on GitHub →Drop-in library, transparent proxy, or MCP server — wire it in however your stack works.
Targets the output that blows up agent context — command output, stack traces, logs.
Shrinks retrieved chunks before they reach the model, leaving room for more context.
Tuned to cut tokens without changing what the model returns.
Open Compress — Product demo from the project's brand account (1k likes). Note: no third-party post cleared 100 likes — the repo is niche on X.
Open on X →Google's NotebookLM is great until you want your own models, your own data boundaries, or features Google doesn't ship. Open Notebook is the open-source implementation — bring your own LLM, organize sources into notebooks, and generate grounded notes and audio overviews on your terms.
View on GitHub →Run it on your own infrastructure — your sources stay under your control.
Plug in the LLM and TTS providers you want instead of being locked to one vendor.
Organize documents, links, and media, then chat and generate grounded notes.
Turn your sources into podcast-style audio like NotebookLM's signature feature.
Better Stack — fresh demo of Open Notebook as a self-hosted, API-driven NotebookLM (26.9k views) · Open on YouTube (7:48 → 3:33) →
AwesomeFOSS — feature + setup walkthrough (3.7k views) · Open on YouTube (8:25) →
I've used Zapier across multiple businesses for over 10 years. They've been doing automation reliably long before this moment — and the MCP service brings that same reliability to any AI agent you're building with.
Try Zapier MCP →Exposes 8,000+ tools as MCP resources any agent can call.
Claude Code, OpenCode, Cursor — anything that speaks MCP.
Enable the tools you want, grab the MCP endpoint, paste it into your agent.
Zapier handles auth, rate limits, and reliability for you.
A Codex skill that generates hand-drawn 16:9 illustrations around a recurring "小黑" (little-black) blob mascot — white background, a few red / orange / blue annotations. @venturetwins fed it the X recommendation-algorithm repo and got instant explainer art.
View on GitHub →Paste any text — narrative, technical, or code — and get matching explainer graphics.
Every image features the same blob mascot for a cohesive look.
Clean, slide-ready frames with light red / orange / blue accent annotations.
Install it into Codex and call it on any document.
@venturetwins — "Stumbled upon a Codex skill that creates cool illustrations to explain topics" — fed it the X algorithm repo (3k likes)
Open on X →Built by Luca Rossi of the Refactoring newsletter to run his own 10,000-note vault. Notes stay plain markdown in a git repo — no accounts, no lock-in. AI-first (works with Claude Code and Codex) but keyboard-first for power users. Brand new — only a handful of stars so far.
View on GitHub →Notes are plain markdown — portable, editor-agnostic, no export step.
Every vault is a git repo: full history, any remote, zero server dependency.
No accounts or subscriptions; stop using it and you lose nothing.
Ships an AGENTS file so Claude Code or Codex can edit the vault directly.
From Riley Brown: type a prompt, an agent writes the code, it goes live in the cloud, and you preview it in-app. Web apps run in Daytona sandboxes; iOS apps compile via Chorus cloud Xcode and install over the air. Built with SwiftUI + Convex + Claude. New — ~20 stars.
View on GitHub →Toggle between polished web apps and real SwiftUI iOS apps.
Build status streams in real time; follow-ups edit the app, compile errors auto-repair.
Ask the chat for the download link and the agent posts an OTA install link.
No auth, no waitlist — runs on your Convex deployment and your API keys.
@rileybrown — "Mythos (Fable) is AGI" — rebuilt the Lovable mobile app in 2 prompts and says his version "SMOKED it" (6.3k likes)
Open on X →Generic AI gives you text; this gives you structure. PM Skills Marketplace encodes proven frameworks from Teresa Torres, Marty Cagan, and Alberto Savoia into agent skills and slash commands. Install once in Claude Code or Cowork and get all 9 plugins — from /discover to /plan-launch.
Discovery, strategy, execution, research, analytics, growth, GTM, toolkit, AI shipping.
/discover runs ideation → assumptions → prioritization → experiments end to end.
OSTs, RICE, Lean Canvas, Porter's Five Forces — the rigor built into daily workflow.
SKILL.md files run in Codex, Cursor, Gemini CLI, and any tool that reads the format.
A reference repo from OpenAI showing how to build Codex plugins: each lives under plugins/<name> with a manifest plus optional skills, MCP, agents, commands, and hooks. The richer examples double as templates you can copy.
use_figma, Code to Canvas, Code Connect, and design-system rules.
build-ios-apps, build-macos-apps, and build-web-apps for SwiftUI, AppKit, and web stacks.
Planning/research, deploys, and programmatic video as ready-made bundles.
.codex-plugin/plugin.json plus skills/, .mcp.json, agents/, commands/, hooks.
@OpenAIDevs — Official launch of role-specific Codex plugins (official account, 1.2k likes)
Open on X →Ask an agent to check Twitter or read a YouTube transcript and it usually chokes on paywalled APIs and 403s. Agent Reach is a scaffolding installer that wires up free upstream tools so any shell-capable agent gets internet access. Each channel is a pluggable file you can swap.
View on GitHub →Paste a URL to your agent and it installs the CLI, deps, and a SKILL.md guide.
Web, YouTube, RSS, GitHub, Twitter, Reddit, Bilibili, XiaoHongShu, LinkedIn, Weibo, and more.
Credentials live in ~/.agent-reach with 600 perms — nothing uploaded.
Each platform is one file calling an upstream tool — don't like it, swap it.
Cosmos 3 is a suite of omnimodal world models that jointly process and generate language, images, video, audio, and action in one Mixture-of-Transformers architecture. A Reasoner handles understanding and planning; a Generator simulates futures and produces robot training data.
View on GitHub →Reasoner for text-out reasoning; Generator for vision / sound / action-out simulation.
Cosmos3-Nano (16B) and Super (64B), including text-to-image and image-to-video variants.
Action modeling, policy prediction, and a DROID manipulation policy out of the box.
Diffusers / Transformers for dev, then vLLM-Omni and NIM for OpenAI-compatible serving.
@IlirAliu_ — A builder's thread on what Physical AI actually takes, with Cosmos in the mix (260 likes)
Open on X →@nvidia — "Cosmos 3 just topped 7 physical AI leaderboards" (official account, 337 likes)
Open on X →NVIDIA — the official Cosmos 3 intro: the open model that thinks, generates, and acts (official channel, 8.1k views) · Open on YouTube (1:51 → 1:02) →
See the world-model demos, model cards, and research reports behind Cosmos 3.
Three of this week's repos got flagged for likely fake star / fork campaigns this week. Open the threads and judge for yourself:
AI coding agents ship generic, samey UIs. Taste Skill is an anti-slop framework of portable SKILL.md files that upgrade what Codex, Cursor, and Claude Code produce — with tunable dials for variance, motion, and density. Install with npx skills add.
npx skills add scans the repo and installs every code and image-gen skill.
DESIGN_VARIANCE, MOTION_INTENSITY, and VISUAL_DENSITY from 1–10.
Minimalist, brutalist, soft / high-end, and a stricter GPT / Codex flavor.
Rules target design intent, so they work across React, Vue, and Svelte.
@LexnLin — Creator — "Introducing images-taste skill" (official account, 1.2k likes)
Open on X →DevsKingdom — before / after: Taste Skill on an AI-built interface (1.1k views) · Open on YouTube (5:22 → 2:33) →
Browse the live examples and changelog for what each skill produces.
container is Apple's official CLI for creating and running OCI-compatible Linux containers as fast, lightweight virtual machines on the Mac. It pulls and pushes from any standard registry and is optimized for Apple silicon on macOS 26 — Apache-2.0 licensed.
View on GitHub →Pull and run images from any registry, and push images you build back out.
Each container runs in its own fast VM using macOS 26 virtualization.
Written in Swift and optimized for M-series Macs.
Download the signed package, run container system start, and you're running.
@timsneath — "A sleeper hit from WWDC — container machines" — a prominent developer's take (9.7k likes)
Open on X →LLMs natively "speak" Markdown, and it's token-efficient. MarkItDown converts PDF, Word, PowerPoint, Excel, images, audio, HTML, and even YouTube URLs into structured Markdown — preserving headings, lists, tables, and links. Built by the AutoGen team.
View on GitHub →PDF, PowerPoint, Word, Excel, images, audio, HTML, CSV/JSON/XML, ZIP, EPub, YouTube.
markitdown file.pdf > out.md, or pipe content straight in.
Keeps headings, lists, tables, and links so analysis tools get clean input.
markitdown-ocr for image OCR; Azure Document Intelligence for higher fidelity.
@mdancho84 — Explainer thread — "Microsoft launches a free Python library that converts ANY document to Markdown" (396 likes)
Open on X →100x Engineers — "Don't upload PDFs to Claude, use this instead" (~39k likes) · Open on YouTube →