Hottest Repos
Understand-Anything — 5,625 stars today (30.6k total). Turns any codebase into an interactive knowledge graph you can search and query. Works with Claude Code, Codex, Cursor, Gemini CLI. The "code comprehension as infrastructure" thesis is winning hard.
codegraph — 3,171 stars today, 18.1k this week (24.7k total). Pre-indexed code knowledge graph that's 100% local. Same thesis as Understand-Anything but focused on reducing token usage and tool calls. Two code-graph tools trending simultaneously = category moment.
openhuman — 15.2k stars this week (27.7k total). Private, self-hosted "personal AI super intelligence" written in Rust. GPL-3.0 licensed. The name says it all — people want AI that's theirs.
CloakBrowser — 6.9k stars this week (20.9k total). Stealth Chromium with source-level fingerprint patches. Drop-in Playwright replacement that passes 30/30 bot detection tests. A sign of how much web scraping infrastructure matters in the agent era.
ai-engineering-from-scratch — 3,167 stars today (18.3k total). Comprehensive AI engineering curriculum. Same author's agentmemory (17.8k stars, +6.4k/week) is also surging — persistent memory for coding agents.
AI & ML
anthropics/knowledge-work-plugins — 1,448 stars today (15.2k total). Official open-source plugins for Claude Cowork. Anthropic building the knowledge worker plugin ecosystem in the open.
CLI-Anything — 4,759 stars this week (40.3k total). HKU research: "Making ALL Software Agent-Native." Wraps any software in a CLI layer that agents can operate. Huge implications for tool-use in agentic pipelines.
microsoft/agent-governance-toolkit — 198 stars today (2.2k total). Policy enforcement, zero-trust identity, execution sandboxing for autonomous agents. Covers all 10 OWASP Agentic Top 10 risks. Enterprise-grade agent safety entering open source.
Kronos — 243 stars today (26k total). Foundation model for financial market language — not NLP on earnings calls, but modeling market microstructure directly. Finance-specific foundation models are a trend to watch.
supertonic — 2.7k stars this week (10.4k total). Lightning-fast on-device multilingual TTS via ONNX. Bindings for Swift, Rust, Python, Go, C#, Flutter, Node, WebGPU. The "TTS everywhere" library.
Developer Tools
gstack — 102k stars. Garry Tan's exact Claude Code setup: 23 opinionated tools acting as CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA. The "celebrity dev setup" as product category.
oh-my-pi — 2.4k stars this week (7.3k total). Terminal coding agent with hash-anchored edits, LSP integration, browser, and subagents. Multi-provider (Anthropic, OpenAI). A serious Claude Code / Codex competitor.
cmux — 598 stars today (19.4k total). Ghostty-based macOS terminal with vertical tabs and notifications designed specifically for AI coding agents. The "AI-native terminal" niche is real.
presenton — 1.7k stars this week (6.9k total). Open-source AI presentation generator and API. Gamma/Beautiful.ai alternative you can self-host.
Cursor plugins — 303 stars today (790 total). Official Cursor plugin specification. Still tiny but signals Cursor building an extensibility ecosystem.
Hidden Gems
dograh — 2.9k stars, +1.1k/week. Open source voice agent platform with telephony, WebRTC, and Pipecat integration. Self-hosted Vapi alternative. If you're building voice agents, this is the one to watch.
firecrawl/pdf-inspector — 1.3k stars. Rust library that intelligently detects scanned vs text-based PDFs for smart OCR routing. From the Firecrawl team. Tiny but solves a real pain point in document processing pipelines.
CADAM — 3.6k stars. Open-source text-to-CAD web app using OpenSCAD + WASM. Describe a part in natural language, get a 3D model. Mechanical engineering meets LLMs.
Notable Trends
Code knowledge graphs are the new RAG. Both Understand-Anything (#1 daily) and codegraph (#1 weekly) are racing to be the default "give your agent pre-indexed codebase context" tool. The shift from "let the agent grep around" to "hand it a structured graph" is happening fast.
The CLAUDE.md / agent skills economy is absurd. ECC (192k stars), superpowers (206k), andrej-karpathy-skills (155k), gstack (102k), taste-skill (19.5k), stop-slop (4.3k), cybersecurity-skills (9.1k). Prompt engineering has become a product category with six-figure star counts. We're watching "dotfiles for AI" become one of the most-starred categories on GitHub.