Hottest Repos
mvanhorn/last30days-skill — 3,558 stars today (34.2k total). Going parabolic. From 28k → 30.5k → 34.2k over three days. The "skills as products" poster child just gained 6k stars in 72 hours. Agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web. 123 open issues — community is building extensions faster than the maintainer can triage. This is what escape velocity looks like.
RyanCodrai/turbovec — 1,730 stars today (8.7k total). Second day, still accelerating. From 6.9k → 8.7k. Rust vector index built on Google's TurboQuant — fits 31 GB of float32 embeddings into 4 GB, no training step, online ingest. Hand-written NEON/AVX-512 kernels. The RAG infrastructure layer's most credible new challenger to FAISS. MIT.
roboflow/supervision — 1,140 stars today (42.3k total). Sudden spike on an established giant. Reusable computer vision toolkit — detection, tracking, segmentation, classification — all composable. Pushed today. At 42k stars and 3.8k forks this is a production staple, not a toy. Something triggered a visibility event. MIT.
Panniantong/Agent-Reach — 796 stars today (23.9k total). Third day, steady climb. One CLI scraping Twitter, Reddit, YouTube, GitHub, Bilibili, XiaoHongShu — zero API fees. Where last30days-skill synthesizes, Agent-Reach provides raw access. The "agent data layer" category now has two 20k+ players. MIT.
santifer/career-ops — 665 stars today (50.4k total). New to digest but somehow 50k stars. AI-powered job search system built on Claude Code — evaluates offers (A-F scoring), generates ATS-optimized PDFs, scans job portals via Playwright, batch-processes with sub-agents. Creator used it to evaluate 740+ listings and land a Head of Applied AI role. Featured in WIRED and Business Insider. The first truly viral "AI for life admin" tool. MIT.
AI & ML
google/skills — 481 stars today (12.3k total). New to digest. Google enters the skills game. Official agent skills for Google Cloud products — Gemini API, BigQuery, Cloud Run, Firebase, GKE, AlloyDB, plus Well-Architected Framework skills for security, reliability, and cost. Installable via
npx skills add google/skills. Google's answer to OpenAI/plugins — but cloud-infrastructure-focused rather than consumer apps. Apache-2.0.MemPalace/mempalace — 237 stars today (54.8k total). Fifth day in digest, decelerating but entrenched. The local-first memory system (96.6% recall@5, zero API calls) is now the clear category winner at 54.8k. Growth slowing from 500+/day to 237 — normal maturation curve.
Andyyyy64/whichllm — 103 stars today (3.4k total). New to digest. One command to find the best local LLM for your hardware. Auto-detects GPU/CPU/RAM, ranks models from HuggingFace by what actually fits and performs.
whichllm --gpu "RTX 4090"shows you the answer. Can simulate GPUs before you buy. Hasupgradeto compare GPU candidates andplanto find what GPU a model needs. The missing link between "I want to run local LLMs" and "which one?" MIT.Anti-slop movement: 83.9k combined — taste-skill (38.2k), impeccable (36.2k), stop-slop (9.5k). Up from 80k just yesterday. impeccable gained 3.6k this week. The "teach your AI taste" movement is now larger than most standalone developer tools.
Developer Tools
refactoringhq/tolaria — 649 stars today (13.4k total). Second day, holding strong. Desktop app for markdown knowledge bases — files-first, git-first, offline-first. AI-native with Claude Code, Codex CLI, and Gemini CLI setup paths. Think Obsidian but fully open-source. Tauri + React. AGPL-3.0.
openai/plugins — 296 stars today (2.2k total). Third day in digest. Official Codex plugin collection. With Google/skills now also live, the "official skills ecosystem" war is heating up: Google (cloud infra), OpenAI (consumer apps + Figma/Notion), and the community (last30days, taste-skill, career-ops). Vendor-neutral skills are winning on star counts by 10x.
aaif-goose/goose — 699 stars today (48k total). Open-source AI agent in Rust — install, execute, edit, test with any LLM. Quiet but massive at 48k. The Rust agent ecosystem (goose + turbovec + sniffnet) is having a good week.
Hidden Gems
harehare/mq — 759 stars, 112 today. jq for Markdown. Query, filter, and transform Markdown files from the command line with a jq-like syntax. Written in Rust. Has a REPL, VSCode extension, LSP support, and even a debugger. Purpose-built for LLM workflows where Markdown is the primary input/output format. Has an online playground. Pushed today. This is the kind of tool you don't know you need until you try it. MIT.
microsoft/mxc — 768 stars, 65 today. The sandbox for AI agent code execution. Microsoft eXecution Container — sandboxed code execution for untrusted model output, plugins, and tools. Cross-platform (Windows/Linux/macOS), multiple containment backends from process sandboxes to full VMs, JSON-based policy configuration, TypeScript SDK. Early preview but actively pushed. If you're building agents that run generated code, this is the missing security layer. MIT.
Notable Trends
The Great Skills Land Grab: Google (google/skills), OpenAI (plugins), and dozens of community repos are racing to define the agent skills ecosystem. Community-built skills (last30days at 34k, taste-skill at 38k, career-ops at 50k) are crushing vendor-official collections (google/skills 12k, openai/plugins 2.2k) on raw adoption. The lesson: vendor-neutral, instantly-useful skills win over platform-locked collections.
AI escaping the IDE: career-ops (job search), MoneyPrinterTurbo (video generation, 82k), Agent-Reach (web scraping) — the fastest-growing AI repos are no longer developer tools. They're life-admin tools, content factories, and data pipelines. The "AI agent for X" template has product-market fit across every vertical, not just coding.