Hottest Repos
NousResearch/hermes-agent — 144,564 stars (+2,229 today). The 144K-star agent just had its biggest day in weeks. Self-improving skills, cross-platform gateway (Telegram/Discord/Slack/WhatsApp/Signal from one process), 200+ model support, serverless hibernation. v0.13.0 is polished enough to explain the surge — this isn't hype, it's adoption. https://github.com/NousResearch/hermes-agent
anthropics/financial-services — 20,181 stars (+1,734 today). Day 6: crossed 20K, settling into cruising altitude. Trajectory: 540 → 1,367 → 3,662 → 3,077 → 1,479 → 1,734. The slight re-acceleration suggests a second wave — likely enterprise teams evaluating after the initial dev rush. 69 open PRs signal real community building. https://github.com/anthropics/financial-services
farion1231/cc-switch — 67,231 stars (+1,364 today). The agent CLI control panel keeps compounding. Unified management for Claude Code, Codex, Gemini CLI, OpenCode, and OpenClaw — 50+ provider presets, MCP server sync across 4 apps, usage dashboards, cloud config sync. At 67K stars this is infrastructure, not a side project. https://github.com/farion1231/cc-switch
CloakHQ/CloakBrowser — 5,973 stars (+1,325 today). Day 3: stealth browser sustains explosive growth. +3.5K in three days. Source-level C++ Chromium patches, 0.9 reCAPTCHA v3 scores, passes Cloudflare Turnstile. The bot-detection arms race continues escalating — now with a
pip installon-ramp. https://github.com/CloakHQ/CloakBrowserHmbown/DeepSeek-TUI — 25,100 stars (+22K this week). The run of the year keeps running. From 1.1K to 25K in twelve days. v0.8.27 just landed 17 community PRs — terminal flicker fixes, desktop notifications, i18n onboarding, MCP hot-reload. This is no longer a breakout; it's a category winner building its ecosystem. https://github.com/Hmbown/DeepSeek-TUI
AI & ML
datawhalechina/hello-agents — 47,547 stars (+1,267 today). Agent education hits escape velocity in China. 16-chapter curriculum from foundations to multi-agent capstone projects, 5.7K forks. Combined with easy-vibe (9.8K, +808 today), Chinese developer education is training an entire generation agent-native. https://github.com/datawhalechina/hello-agents
TauricResearch/TradingAgents — 73,791 stars (+8,872 this week). Twelfth consecutive week of massive growth. The finance-AI gravity well is inescapable. AI-Trader (16.1K, +845 today) and dexter (25.3K) continue orbiting alongside it. Three 15K+ finance agent repos trending simultaneously is now the norm, not the exception. https://github.com/TauricResearch/TradingAgents
VectifyAI/PageIndex — 30,640 stars (+4,328 this week). Vectorless RAG crossed 30K. Hierarchical tree search instead of embeddings — 98.7% accuracy on FinanceBench. No vector DB, no chunking, reasoning-driven retrieval that mimics how domain experts actually read documents. The "similarity ≠ relevance" thesis is winning. https://github.com/VectifyAI/PageIndex
HKUDS/AI-Trader — 16,115 stars (+845 today). Accelerating again — up from +695 two days ago. Agent-native trading with copy-trade sync across brokers, paper trading with real Polymarket data. The "agents as market participants" category is now too big to ignore. https://github.com/HKUDS/AI-Trader
heygen-com/hyperframes — 17,223 stars (+401 today). Write HTML. Render video. Deterministic HTML-to-MP4 via headless Chrome + FFmpeg. 50+ pre-built components, agent-optimized CLI, Apache 2.0 (unlike Remotion). The pitch: agents already know HTML, so video generation is just another rendering target. https://github.com/heygen-com/hyperframes
Developer Tools
bytedance/UI-TARS-desktop — 32,927 stars (+956 today). ByteDance's vision-to-action agent stack re-enters daily top 3. Connecting multimodal AI to desktop automation — the "see screen, click things" paradigm is getting serious enterprise backing. +2.2K this week. https://github.com/bytedance/UI-TARS-desktop
decolua/9router — 8,208 stars (+942 today). Third week, crossed 8K. Free AI coding router for Claude Code, Codex, Cursor — 40+ providers, auto-fallback, token compression. Weekly gains steady at 3.3K. The "route everything through free tiers" play has proven durability. https://github.com/decolua/9router
rohitg00/agentmemory — 4,633 stars (+604 today). Persistent memory for coding agents sustains momentum. Background server auto-captures actions, consolidates into searchable memory, injects relevant context. The agent memory layer that works across different agent CLIs. https://github.com/rohitg00/agentmemory
millionco/react-doctor — 7,962 stars (+340 today). "Your agent writes bad React. This catches it." Health score 0-100 scanning state management, performance, security, accessibility, dead code. Works with 50+ coding agents, GitHub Actions integration. The quality-gate-for-AI-code category is real now. https://github.com/millionco/react-doctor
docusealco/docuseal — 16,359 stars (+4,020 this week). Open-source DocuSign alternative surges. v2.5.3 just shipped. 152 releases, PDF form builders, 12 field types, self-hosted. The enterprise OSS document signing niche is heating up. https://github.com/docusealco/docuseal
Hidden Gems
tinyhumansai/openhuman — 1,339 stars (+501 today). Personal AI with 118-service OAuth integration. Connects Gmail, Notion, GitHub, Slack — auto-fetches every 20 minutes, builds a Memory Tree in local SQLite + Obsidian-compatible vault. TokenJuice compression cuts 80% of token waste. Animated mascot joins your Google Meet calls. Rust + TypeScript, fully local-first. The "AI assistant that actually knows your context" space has a serious new contender. https://github.com/tinyhumansai/openhuman
cartesiancs/map3d — 715 stars (+165 today). Generate 3D city maps from OpenStreetMap with React-Three-Fiber. Buildings, roads, GLB export — all free. 39 commits, MIT licensed, active roadmap (textures, heightmaps next). Tiny but genuinely useful for game devs, architects, and urban planners. https://github.com/cartesiancs/map3d
RhysSullivan/executor — 1,624 stars (+37 today). The missing integration layer for AI agents. Unified tool catalog across OpenAPI, GraphQL, MCP, and custom JS — agents discover tools via semantic search, execute with centralized auth. Acts as MCP server for Cursor/Claude Code/OpenCode. Solves the "every agent needs its own credential management" problem. https://github.com/RhysSullivan/executor
Notable Trends
Agent quality gates are now a category. react-doctor (8K), agentmemory (4.6K), and the agent-skills repos (mattpocock 72K, addyosmani 39.4K) all address the same insight: AI-generated code needs specialized review tooling. The market has moved past "agents that write code" to "systems that ensure agent code is good." This is the maturing signal.
The skills/agent-skills format war is over — everyone won. mattpocock/skills (+12.7K this week to 72K), addyosmani/agent-skills (+10.7K to 39.4K), huggingface/skills (10.5K), and now wanshuiyin/Auto-claude-code-research-in-sleep (8.9K) — four different approaches to "how to instruct agents" all growing simultaneously. The category is big enough for multiple winners, and the winner is the
.claude/skills/directory pattern itself.