Hottest Repos
obra/superpowers — 199,904 stars (+1,776 today). 200K TODAY. Less than 100 stars away. The agentic skills framework becomes the first open-source dev tool to hit 200K stars. 17.8K forks. This is a milestone for the entire ecosystem — the "Linux kernel of agent skills" is now one of the most-starred repos on all of GitHub. https://github.com/obra/superpowers
tinyhumansai/openhuman — 23,488 stars (+3,603 today, +17,793 this week). Day 10: nearly 10x from launch. From 2.5K to 23.5K in ten days. 2,100 forks, 207 open issues. Another 3.6K-star day — still not decelerating. The Rust-powered personal AI agent is May's defining repo and is on pace for 30K by weekend. https://github.com/tinyhumansai/openhuman
mattpocock/skills — 96,785 stars (+19,038 this week). 100K in ~1 day. Jumped from 94K yesterday to 97K. Eight straight weeks of 3K+/day. The fastest project ever from 50K→100K — and it's a
.claudedirectory. https://github.com/mattpocock/skillscolbymchenry/codegraph — 9,148 stars (+1,910 today, +4,650 this week). Day 7: nearly 4x from launch. From 2.4K to 9.1K. Pre-indexed code knowledge graph for Claude Code/Codex/Cursor. The "give agents a map" thesis just hit escape velocity — nearly doubled since yesterday's 6.4K. https://github.com/colbymchenry/codegraph
multica-ai/andrej-karpathy-skills — 140,558 stars (+2,620 today). The single CLAUDE.md that 140K developers use. Derived from Karpathy's observations on LLM coding pitfalls. 14.4K forks. The most-starred single-file project in the agent skills category. https://github.com/multica-ai/andrej-karpathy-skills
AI & ML
HKUDS/ViMax — 5,892 stars (+692 today). Day 3: agentic video generation accelerating. Director, Screenwriter, Producer, and Video Generator all-in-one — input a concept, get a complete video. Same HKU team behind CLI-Anything (38K stars). Idea2Video and Novel2Video modes. MIT licensed. https://github.com/HKUDS/ViMax
ruvnet/RuView — 61,779 stars (+8,076 this week). WiFi-as-radar is real. Turns commodity WiFi signals into spatial intelligence — presence detection, breathing/heart rate, pose estimation, fall detection — through walls, no cameras. $9 ESP32 sensor, 8KB quantized model, 100% presence accuracy on validation set. 105-module edge catalog. The spec sheet reads like science fiction but the 1,463 passing tests say otherwise. https://github.com/ruvnet/RuView
rohitg00/agentmemory — 14,892 stars (+1,121 today, +8,390 this week). Week 3: the "agents need memory" category is cemented. 4x in two weeks. Background capture, auto-consolidation, cross-session context injection. https://github.com/rohitg00/agentmemory
BigBodyCobain/Shadowbroker — 8,458 stars (+320 today). OSINT platform crossed 8K. 60+ real-time feeds on one map with an HMAC-signed agentic command channel. Quiet, steady growth — this has legs. https://github.com/BigBodyCobain/Shadowbroker
HKUDS/CLI-Anything — 38,376 stars (+930 today). HKU's agent-native CLI wrappers still climbing. 50+ desktop apps (Blender, Obsidian, QGIS, Godot) exposed as CLIs for agents. 3.5K forks. https://github.com/HKUDS/CLI-Anything
Developer Tools
anthropics/claude-plugins-official — 20,706 stars (+706 today). NEW: Anthropic's official plugin marketplace. Curated directory of Claude Code plugins — internal and third-party. 680 open issues (mostly plugin submissions). The
/plugin installcommand is now a first-class distribution channel. https://github.com/anthropics/claude-plugins-officialmillionco/react-doctor — 10,446 stars (+1,796 this week). "Your agent writes bad React. This catches it." From the Million.js team.
npx react-doctorgives a 0-100 health score with actionable diagnostics — state/effects, performance, architecture, security, accessibility. Works as a Claude Code skill to teach agents React best practices before they write bad code. The "agent quality guard" pattern is becoming a category. https://github.com/millionco/react-doctormsitarzewski/agency-agents — 102,740 stars (+1,714 today). Crossed 100K. Complete AI agency in a box — specialized agent roles (frontend, community, copy, QA) with personality and processes. 16.9K forks. https://github.com/msitarzewski/agency-agents
rtk-ai/rtk — 51,739 stars (+1,214 today). Past 51K. The Rust CLI proxy reducing LLM token consumption 60-90%. Essential agent infrastructure, quiet growth. https://github.com/rtk-ai/rtk
can1357/oh-my-pi — 5,285 stars (+237 today). The maximalist terminal agent. Fork of Pi — 40+ providers, 32 built-in tools, 13 LSP ops, 27 DAP ops, ~27K lines of Rust core. Hash-anchored edits, subagents, browser integration. One to watch. https://github.com/can1357/oh-my-pi
Hidden Gems
zakirullin/files.md — 2,164 stars (+468 today). A dead-simple, local-first markdown notes app — no cloud, no build system, just open
index.html. Go sync server optional. Five years of patient development, now viral. The codebase is intentionally small enough for "one person or an LLM to fit the whole project in head." The anti-Notion. https://github.com/zakirullin/files.mdArmur-Ai/Pentest-Swarm-AI — 1,208 stars. Autonomous pentesting using a swarm of AI agents (Go + Claude API). Orchestrates recon, classification, exploitation, and reporting specialists with ReAct reasoning. Bug bounty, continuous monitoring, and CTF modes. AGPL-3.0. The security-AI intersection is still underexplored — this is one of the better implementations. https://github.com/Armur-Ai/Pentest-Swarm-AI
git-ai-project/git-ai — 1,919 stars. Rust-based Git extension that tracks which code in your repo was AI-generated. As AI-written code becomes the majority in many repos, provenance tracking shifts from nice-to-have to necessity. Small, practical, underappreciated. https://github.com/git-ai-project/git-ai
Notable Trends
Two historic milestones today/tomorrow. obra/superpowers crosses 200K stars (literally hours away) and mattpocock/skills crosses 100K (~1 day). Both are
.claudeskill collections. The agent skills category now has two repos in the top 100 most-starred projects on all of GitHub. This would have been unthinkable six months ago.The "agent quality guard" pattern is emerging as its own category. react-doctor (catches bad React from agents), codegraph (gives agents a code map to reduce errors), andrej-karpathy-skills (teaches agents to avoid pitfalls), Imbad0202/academic-research-skills (catches hallucinated citations). The meta-game has shifted from "make agents do more" to "make agents do it right." The most successful new repos are the ones that constrain and guide agent behavior rather than just expand it.