Hottest Repos
tinyhumansai/openhuman — 10,527 stars (+1,601 today). Day 6: crossed 10K. From 2.5K to 10.5K in one week — a 4x. The "personal AI super intelligence" with 118+ OAuth integrations and Memory Tree just validated its thesis. Rust-powered, Obsidian-compatible. The velocity is still accelerating. https://github.com/tinyhumansai/openhuman
CloakHQ/CloakBrowser — 12,485 stars (+9,120 this week). Day 9: tripled from 3.9K. The stealth Chromium that passes every bot detection test is now firmly a five-figure project. Source-level C++ patches, drop-in Playwright replacement. The fastest weekly gain on the board. https://github.com/CloakHQ/CloakBrowser
mattpocock/skills — 86,706 stars (+18,278 this week). Sixth straight week of massive growth. Gained 2K today alone. From 68K to 87K this month. At this pace it crosses 100K by end of May. The
.claudeskills collection is infrastructure now, not a project. https://github.com/mattpocock/skillsobra/superpowers — 193,832 stars (+1,281 today). Approaching 200K. Jesse Vincent's agentic skills framework is 6K away from a milestone no open-source dev tool has ever hit. 16.8K forks. Clockwork-steady 1,200-1,800/day. https://github.com/obra/superpowers
ruvnet/RuView — 58,206 stars (+990 today, +5K this week). WiFi-as-radar refuses to plateau. Turns ESP32 WiFi CSI into spatial intelligence and vital signs. No cameras. Approaching 60K with no sign of slowing. https://github.com/ruvnet/RuView
AI & ML
anthropics/financial-services — 23,751 stars (+9,480 this week). Anthropic's vertical play keeps compounding. Reference agents for pitch decks, equity research, GL reconciliation, KYC screening. Partners with LSEG and S&P Global. This is how model vendors capture enterprise value — open-source the wedge, monetize the depth. https://github.com/anthropics/financial-services
rohitg00/agentmemory — 10,194 stars (+6,865 this week). Crossed 10K. From 3.4K to 10K in 8 days. Persistent memory for coding agents — background capture, auto-consolidation, context injection. The "agents need memory" thesis is now undeniable. https://github.com/rohitg00/agentmemory
Hmbown/DeepSeek-TUI — 30,475 stars (+8,701 this week). DeepSeek's answer to Claude Code hits 30K. Rust TUI with 1M-token context, streaming reasoning, approval gates, auto mode. Every major model family now has a terminal agent. https://github.com/Hmbown/DeepSeek-TUI
supertone-inc/supertonic — 6,711 stars (+745 today). On-device TTS crossed the utility threshold. 31-language, lightning-fast ONNX Runtime inference. SDKs for 7 languages. Voice Builder for permanent clones. Edge-AI audio is real. https://github.com/supertone-inc/supertonic
K-Dense-AI/scientific-agent-skills — 23,027 stars (+669 today). 135 scientific agent skills. Genomics, drug discovery, materials science, clinical research — plus 100+ database connectors. The deepest vertical skill pack on GitHub. https://github.com/K-Dense-AI/scientific-agent-skills
Developer Tools
millionco/react-doctor — 9,750 stars (+3,314 this week). NEW: "Your agent writes bad React. This catches it." One command → 0-100 health score with actionable diagnostics. Installs as a skill for Claude Code, Cursor, Codex. Catches state & effects, performance, architecture, security, a11y issues. Turns agent code review into a linting step. https://github.com/millionco/react-doctor
colbymchenry/codegraph — 2,376 stars (+397 today). NEW: Pre-indexed code knowledge graph for Claude Code. Claims 92% fewer tool calls, 71% faster exploration. Builds a symbol graph so Explore agents query relationships instead of scanning files. Benchmarks across VS Code, Excalidraw, Swift compiler look impressive. https://github.com/colbymchenry/codegraph
decolua/9router — 11,056 stars (+5,377 this week). AI model router crossed 11K. Smart proxy connecting Claude Code/Codex/Cursor to 40+ providers with 3-tier auto-fallback. RTK token compression saves 20-40% input tokens. 489 open issues = explosive adoption pain. https://github.com/decolua/9router
AlexsJones/llmfit — 26,247 stars (+211 today). "What runs on my hardware?" One command scores hundreds of models on quality, speed, fit, and context for your specific GPU/RAM. Community leaderboard with real tok/s from actual users. The practical answer to "which local model should I use." https://github.com/AlexsJones/llmfit
Hidden Gems
dograh-hq/dograh — 1,298 stars (+297 today). Open-source alternative to Vapi & Retell for production voice agents. Drag-and-drop workflow builder, self-hostable, BSD-licensed. Built by YC alumni. Pipecat-based with WebRTC, multi-provider TTS/STT. The "build your own Bland.ai" starter. https://github.com/dograh-hq/dograh
NVIDIA-AI-Blueprints/video-search-and-summarization — 1,293 stars (+273 today). GPU-accelerated vision agents for video analytics — reference architectures from NVIDIA with VLM and RAG support. The blueprint for building "ask questions about video" products. https://github.com/NVIDIA-AI-Blueprints/video-search-and-summarization
Raiper34/spooty — 2,246 stars (+184 today). Self-hosted Spotify downloader in TypeScript. Simple, MIT-licensed, and apparently quite popular this week. The perennial demand for music ownership meets modern self-hosting. https://github.com/Raiper34/spooty
Notable Trends
Agent skills are eating software. The top 5 trending repos by weekly growth include 3 skills-related projects (mattpocock/skills 87K, react-doctor 10K, academic-research-skills 8K). Every tool is becoming a
.claudedirectory entry. The category went from zero to dominant in 3 months.Every model family now has a terminal agent. DeepSeek-TUI (30K) joins Claude Code (161K opencode), Codex, Copilot CLI, and Gemini CLI. The "coding agent in your terminal" form factor is settled — the competition is now about model quality and context handling, not UX innovation.