Hottest Repos
Graphify-Labs/graphify — Turn any codebase into a queryable knowledge graph. +1,028 today (84.5k total). New to digest. YC S26. Uses tree-sitter AST parsing — no LLM, fully local for code. Not a vector index — a real traversable graph. Works with Claude Code, Codex, Cursor, Gemini CLI, and 15+ others. One command:
/graphify .→ interactive graph.html + GRAPH_REPORT.md. 8,333 forks. Knowledge graphs are eating RAG. https://github.com/Graphify-Labs/graphifyOpenCut-app/OpenCut — Open-source CapCut alternative. +1,077 today (65.8k total). New to digest. Being rewritten from scratch: Rust core, plugin-first architecture, MCP server for AI agents, headless mode for batch rendering, desktop+mobile+browser from one codebase. The classic version still runs at opencut.app. 6,959 forks, 333 open issues. The open-source video editor space just got serious. https://github.com/OpenCut-app/OpenCut
HKUDS/Vibe-Trading — Multi-agent trading platform. +1,148 today (21.6k total). Day 2, growth accelerating (+776→+1,148). Now the #1 daily Python trending repo. The open-source trading agent category is consolidating around this — TradingAgents (92.8k) is the giant, but Vibe-Trading has the momentum. https://github.com/HKUDS/Vibe-Trading
pbakaus/impeccable — Design language for AI coding agents. +2,272 this week (46.3k total). New to digest. 23 commands (
/impeccable audit,polish,critique,bolder,quieter...), 46 deterministic detector rules that run with no LLM. Evolved from Anthropic'sfrontend-designskill. The anti-AI-slop category now has two serious players: impeccable (46k, comprehensive) vs hallmark (5k, opinionated). impeccable is the framework; hallmark is the taste filter. https://github.com/pbakaus/impeccableNutlope/hallmark — Anti-AI-slop design skill. +802 today (5k total). Day 2, daily stars quadrupled (+210→+802). About to cross 5k. Momentum is real — developers are clearly tired of every AI-generated UI looking the same. https://github.com/Nutlope/hallmark
AI & ML
OfficeCLI continues its tear: +6,978 this week (16k total). The biggest weekly gainer by absolute stars. Agent-readable Office docs was an unmet need — now it's infrastructure. https://github.com/iOfficeAI/OfficeCLI
claude-video crossed 8k: +4,353 this week. Day 6+ in digest, still accelerating. Video comprehension for agents is clearly a must-have category. https://github.com/bradautomates/claude-video
OmniRoute at 16.8k (+4,506/week), CubeSandbox at 10k (+2,490/week), strix at 41.2k (+4,143/week). The agent infrastructure trio — routing, sandboxing, and security — all holding their positions. The plumbing layer is solidifying.
Gentleman-Programming/engram — Persistent memory for AI coding agents. +242 today (5.3k total). New to digest. Single Go binary, SQLite + FTS5, MCP server, CLI, HTTP API, and TUI. Works with every major agent. One-liner setup per agent. The "agent amnesia" problem finally has a clean, agent-agnostic solution. https://github.com/Gentleman-Programming/engram
Developer Tools
orca pulls ahead of herdr: orca at 18.1k (+4,481/week) vs herdr at 16.1k (+3,928/week). Week 3 for both. Orca's GUI+mobile approach is winning the star race, but herdr's terminal-native philosophy still has a loyal following. The "agent fleet management" category is a real market now.
facebook/astryx — Meta's agent-ready design system. +2,397 this week (8.6k total). Fully customizable, built for AI coding agents to consume and generate against. Meta is positioning this as the default component library for agent-built apps. https://github.com/facebook/astryx
ColeMurray/background-agents — Open-source Ramp Inspect clone. +233 today (2.4k total). New to digest. Background coding agents with web UI, Slack, GitHub PR, and Linear integration. Multiplayer sessions, parallel sub-tasks in separate sandboxes, cron/webhook triggers. Self-hosted, single-tenant. If you want agents working overnight without watching them, this is the open-source play. https://github.com/ColeMurray/background-agents
github/spec-kit — GitHub's Spec-Driven Development toolkit. +508 today (120.5k total). Steady growth at massive scale. GitHub is betting that the future of software development is writing specs and letting agents build — spec-kit is their official starter kit for that workflow. https://github.com/github/spec-kit
Hidden Gems
raine/claude-code-proxy — Use Claude Code's harness with your ChatGPT, Kimi, Cursor, or Grok subscription. 283 stars, +32 today. Rust binary, local Anthropic-compatible proxy. Born from frustration with Anthropic's tightening usage limits while OpenAI remains generous. If you love Claude Code's UX but want to use other models, this is it. https://github.com/raine/claude-code-proxy
Vexa-ai/vexa — Self-hosted meeting transcription API for Google Meet, Teams, and Zoom. 2,510 stars, +72 today. Sends a real bot into calls, streams speaker-attributed transcripts via WebSocket, compiles meetings into Markdown in a git repo. Sandboxed agents process the knowledge base. Apache-2.0, air-gap-ready. The meetily competitor that's an API-first platform, not a desktop app. https://github.com/Vexa-ai/vexa
perber/leafwiki — Self-hosted wiki in a single Go binary. 766 stars, +20 today. SQLite + Markdown on disk. No Node.js, no Redis, no Postgres. If you've looked at Wiki.js or Outline and thought "too much to operate" — this is the answer. Docker one-liner, backup is
cp -r. https://github.com/perber/leafwiki
Notable Trends
The Anti-Slop Movement is real. Three design-quality tools are now trending simultaneously: impeccable (46k), hallmark (5k), and Anthropic's own frontend-design skill. Developers are sick of AI-generated UIs that all look identical. The market is splitting into "comprehensive framework" (impeccable) vs "opinionated taste filter" (hallmark) — both are growing fast. Expect every serious AI coding team to adopt at least one by end of month.
"Agent memory" is becoming a commodity. engram (persistent memory via SQLite), Graphify (knowledge graphs), TencentDB-Agent-Memory (4-tier progressive pipeline) — three completely different architectures for the same problem: agents forget everything between sessions. The winner will be the one that works with the most agents with the least setup friction. Right now engram is winning that race with one-liner setup for 12+ agents.