GitHub Trending Weekly Digest — July 27 – August 1, 2026

Another week, another wave of repos climbing the charts. I pulled the daily GitHub Trending Top 5 from July 27 through August 1, merged everything, dropped the duplicates, and ranked what's left by one simple signal: how many days a project actually stayed on the board. Persistence beats a one-day spike — a repo that holds for four days straight is telling you something a viral Tuesday can't.

Seventeen unique projects made the cut this week. Two of them refused to leave, a solid batch bounced around for a couple of days, and the rest each had their single moment in the sun. If there's one word for the week, it's agents — the tooling around AI coding agents is having a real moment.

---

🏆 Chart-Toppers (Multi-Day Streaks)

These are the repos that didn't just trend — they lived on the board.

AIRI — moeru-ai/airi

🔗 github.com/moeru-ai/airi · 4 days on the board 👑

What it does: An open-source, self-hosted AI companion framework built for real-time voice chat. Think of it as an open answer to Neuro-sama-style AI VTubers — except this one runs on your machine, with a Live2D or VRM avatar, a memory system, and the ability to actually play games like Minecraft and Factorio on its own.

Why it matters: Platforms like Character.AI are closed services — they go offline when the stream ends, and your data was never yours to begin with. AIRI flips that: "you own your own digital life." It's not just a chatbot; it sees, remembers, and acts, and it wires up to 30-plus LLM providers so you're never locked into one backend.

Tech: Vue.js + TypeScript with Three.js for the avatar; browser-native WebGPU / WebAudio / WebAssembly for local inference; native inference through Hugging Face's Candle (CUDA/Metal); the xsAI layer abstracts LLM providers; Electron on desktop, Capacitor + PWA on mobile; DuckDB WASM and PostgreSQL + pgvector for storage.

ECC — affaan-m/ECC

🔗 github.com/affaan-m/ECC · 3 days on the board

What it does: A performance-optimization harness for AI coding agents — Claude Code, Codex, Cursor, and friends. It ships 67 specialized agents, 281 reusable skills, and 94 commands, plus built-in testing, review, security scanning, and cross-session learning.

Why it matters: A single LLM can write code, but it has no engineering discipline. Plans get buried in chat history, context management goes chaotic, quality checks depend on you remembering to ask, and whatever it learned this session evaporates by the next. ECC hard-codes a pipeline — plan → test → implement → review → verify → remember → improve — around one core idea: "optimize the context window, persist everything else." Install once, use forever, no re-prompting the whole methodology every time.

Tech: Node.js / TypeScript for scripts and hooks, Python for tooling and a Tkinter dashboard, Bash for system glue; language-specific review rules for Go/Java/C++/Rust/PHP/Perl; and AgentShield, a security auditor backed by 1,282 tests. Compatible with Claude Code, Codex, Cursor, OpenCode, Copilot, Gemini, and Zed. MIT licensed.

---

🔥 Multi-Day Risers (Two Days Each)

A strong middle tier — repos that caught fire and held for a second day.

AI-For-Beginners — microsoft/AI-For-Beginners

🔗 github.com/microsoft/AI-For-Beginners

What it does: Microsoft's open AI curriculum — 12 weeks, 24 lessons, with runnable Jupyter notebooks, quizzes, and hands-on labs.

Why it matters: Plenty of people want to learn AI and have no idea where to start. This breaks the intimidating stuff — neural nets, computer vision, NLP — into small, hands-on lessons so a complete beginner can work through it systematically. Classic democratization of hard knowledge, and it saw the single biggest one-day star jump of the week (+1,592 on August 1).

Tech: Python + Jupyter Notebook; TensorFlow, PyTorch, and Keras; covers CNNs, RNNs, Transformers, GANs, and reinforcement learning, with translations in 50-plus languages.

openwork — different-ai/openwork

🔗 github.com/different-ai/openwork

What it does: A free, open-source desktop app for sharing and reusing AI workflows — positioned as an open-source alternative to Claude Cowork and Codex. Runs on macOS, Windows, and Linux.

Why it matters: The AI tooling ecosystem is fragmented — everyone's on their own island, and workflows don't travel across tools or teammates. openwork lets one set of skills work across multiple agents, and its OpenWork Den control plane gives teams a central place to manage permissions, publish skills, and collaborate.

Tech: Electron + a pnpm-workspace monorepo, built with Vite and Turbo; it uses MCP (Model Context Protocol) to plug into Claude Code, Cursor, Codex, and other agents.

awesome-systematic-trading — paperswithbacktest/awesome-systematic-trading

🔗 github.com/paperswithbacktest/awesome-systematic-trading

What it does: A curated mega-list for systematic (quant) trading — libraries, strategies, books, and courses, all in one place. 97-plus libraries, 40-plus documented strategies, 55-plus books.

Why it matters: Quant resources are scattered all over the web and painful to hunt down. This repo collects and categorizes them so you can find the right backtester, bot, or reading list in one search instead of ten.

Tech: Python-first (NumPy, Pandas, Numba for speed), with some C++/Rust/Go/TypeScript; backtesting frameworks like vnpy, backtrader, vectorbt, and pysystemtrade; trading bots like Freqtrade and Jesse.

jcode — 1jehuang/jcode

🔗 github.com/1jehuang/jcode

What it does: A terminal AI coding agent written in Rust, going head-to-head with Claude Code and Copilot CLI — with an obsessive focus on being the most RAM-efficient of the bunch.

Why it matters: Existing AI coding tools are resource hogs. jcode uses 27.8 MB per session (with local embeddings off) versus 140–386 MB for rivals, and renders its first frame in 14 ms against 383–3,436 ms elsewhere. If you're running several agent sessions in parallel — swarm mode — that gap stops being a benchmark bragging point and starts being the difference between "works" and "doesn't."

Tech: Rust throughout, with a homegrown terminal UI called Handterm (native scrolling, 1000-plus FPS) and its own mermaid renderer; a client/server split for persistent sessions; a semantic memory system that embeds conversations into vectors for context retrieval; MCP support; compatibility with Claude, OpenAI, Gemini, Copilot, and Azure; built-in Firefox automation; and — wild detail — the agent can modify and recompile its own source.

GeoLibre — opengeos/GeoLibre

🔗 github.com/opengeos/GeoLibre

What it does: A free, lightweight, cloud-native open-source GIS platform. Visualize, explore, and analyze geospatial data in the browser, on desktop, on your phone, or inside a Jupyter notebook.

Why it matters: Traditional GIS software (looking at you, ArcGIS) means heavy installs, subscriptions, fiddly config, and often uploading your data to someone else's cloud. GeoLibre strips all that away — open a browser, get 700-plus geoprocessing tools, and your data stays local for privacy. One feature set, every platform.

Tech: React + TypeScript on the front end; MapLibre GL JS for maps; DuckDB-WASM Spatial for in-browser SQL spatial analysis; deck.gl for 3D; Tauri v2 to package cross-platform desktop apps, plus native Android support.

speech-to-speech — huggingface/speech-to-speech

🔗 github.com/huggingface/speech-to-speech

What it does: A fully local voice agent framework from Hugging Face. It implements the whole conversational pipeline — VAD → STT → LLM → TTS — with a modular architecture, and exposes it as an OpenAI Realtime-compatible WebSocket API.

Why it matters: You want a voice assistant but don't want to be chained to a proprietary API. This runs the entire chain on your own hardware, and every stage is swappable, so you get privacy and deep customization. It's already driving thousands of Reachy Mini robots in production.

Tech: Python; Silero VAD v5 for voice detection, Parakeet TDT for speech-to-text (swap in Whisper/Paraformer), an OpenAI-compatible backend for the LLM (self-host with vLLM or llama.cpp), and Qwen3-TTS for speech (or Kokoro, etc.). Runs on CUDA, CPU, or Apple Silicon (MLX).

last30days-skill — mvanhorn/last30days-skill

🔗 github.com/mvanhorn/last30days-skill

What it does: An AI-agent research skill that surveys a topic across social platforms and rolls the results into a grounded, sourced summary. Plugs into Claude Code, Cursor, Copilot, and 50-plus other AI hosts.

Why it matters: The author calls it "a search engine ranked by how real people vote, like, and put money down." Traditional search leans on editorial curation and SEO relevance, which distorts things. This ranks by actual human engagement instead — Reddit upvotes, X likes, YouTube transcripts, TikTok heat, Polymarket odds where people bet real money. Perfect for the "look someone up before you meet them" use case.

Tech: Python 3.12+; yt-dlp for video transcription, a built-in Bird client (Node.js) to search X, and the ScrapeCreators API; parallel multi-source queries with engagement-weighted scoring and cross-source clustering; SQLite for trend storage; Markdown / JSON / HTML output. MIT licensed, 2,700-plus tests.

reverse-skill — zhaoxuya520/reverse-skill

🔗 github.com/zhaoxuya520/reverse-skill

What it does: A cybersecurity skill-routing pack for AI agents (Claude Code, Cursor, Cline). Hand it an APK, a binary, obfuscated JS, a CTF, or a pentest target, and it auto-matches the right methodology and tools.

Why it matters: When an AI does security work, it often just doesn't know which tool to reach for or what order to work in. This freezes scattered tools and hard-won experience into reusable playbooks — 40-plus sub-skills spanning reverse engineering, pentesting, CTF, and firmware analysis — so the agent stops re-learning the same lessons.

Tech: Multi-language integration across Python / Node.js / PowerShell / Bash / Java; drives IDA Pro, radare2, Ghidra, jadx, apktool, and Frida; a master routing script, an MCP server, and a self-evolving knowledge base; Docker support across Windows / Linux / Kali / macOS.

---

⚡ One-Day Wonders

Each of these grabbed a single day on the board — no less interesting, just less sticky.

bitchat — permissionlesstech/bitchat

🔗 github.com/permissionlesstech/bitchat

What it does: A decentralized peer-to-peer chat tool — no account, no phone number, no central server. It moves messages over a Bluetooth mesh network plus the Nostr protocol.

Why it matters: It keeps you talking when the usual network infrastructure is down or can't be trusted — disasters, protests, off-grid areas. Bluetooth mesh handles local, offline multi-hop relay; Nostr adds global reach; the Noise protocol provides end-to-end encryption; and a "triple-tap to wipe" panic button clears your data fast. It posted the biggest single-day star count of the whole week (+1,166).

Tech: Swift (native iOS/macOS), Bluetooth LE, the Nostr protocol, Noise Protocol with XChaCha20-Poly1305 encryption, and LZ4 compression.

Pascal Editor — pascalorg/editor

🔗 github.com/pascalorg/editor

What it does: A browser-based 3D architecture design app — build walls, lay floors, add roofs, place furniture, and share the whole 3D project.

Why it matters: Traditional 3D architectural viz software is expensive and desktop-only. Pascal moves the modeling into the browser with a modern UI and real-time collaboration, lowering the barrier for everyone.

Tech: React 19 + Next.js 16; Three.js (WebGPU) + React Three Fiber for rendering; Zustand + Zundo for state and undo/redo; Zod for validation; Turborepo + Bun for builds; three-bvh-csg for CSG boolean ops; scene data persisted to IndexedDB.

MediaCrawler — NanmiCoder/MediaCrawler

🔗 github.com/NanmiCoder/MediaCrawler

What it does: A multi-platform social-media scraper that pulls public content from Xiaohongshu, Douyin, Kuaishou, Bilibili, Weibo, Baidu Tieba, Zhihu, and more.

Why it matters: Traditional scrapers have to reverse-engineer each platform's JS signing algorithm — hard, and it breaks constantly. MediaCrawler takes a different path: browser automation plus a live login state to grab already-signed requests, sidestepping the whole decryption problem. It supports keyword search, specific post IDs, nested comments, creator pages, proxy pools, and comment word clouds.

Tech: Python (UV package manager), Playwright automation, CDP to attach to an existing Chrome, FastAPI/Uvicorn for the API and WebUI, a Vite front end, with storage to CSV/JSON/SQLite/MySQL.

aisuite — andrewyng/aisuite

🔗 github.com/andrewyng/aisuite

What it does: A lightweight Python library from Andrew Ng's team offering a unified interface across LLM providers — change one model string to switch backends.

Why it matters: Every provider's SDK is different, so wiring up each new one is a chore. aisuite lets you write the code once and run it anywhere, killing vendor lock-in.

Tech: Python; supports OpenAI, Anthropic, Google, Mistral, AWS, Cohere, Ollama, OpenRouter, and more; a provider-adapter architecture; tool calling with auto-generated schemas; built-in file/git/shell toolkits; MCP support; persistent session state. MIT licensed.

Jenkins — jenkinsci/jenkins

🔗 github.com/jenkinsci/jenkins

What it does: The veteran open-source automation server — automating builds, tests, static analysis, and deployment, and all the other repetitive dev chores.

Why it matters: It frees developers from the tedious, repeatable work so they can focus on what actually matters — catching bugs early, keeping code quality up, and shipping faster. Still trending after all these years.

Tech: Java + Maven; a Node.js front-end stack (Yarn, Webpack, Babel, ESLint) plus PostCSS; a modular architecture with 2,000-plus plugins; shipped as a WAR, a Docker image, or native installers; with both weekly and LTS release lines.

superfile — yorukot/superfile

🔗 github.com/yorukot/superfile

What it does: A good-looking, modern terminal file manager — managing files on the command line without it feeling like punishment.

Why it matters: Traditional terminal file managers are ugly and clunky. superfile pairs a pretty TUI with a keyboard-driven workflow to make command-line file ops feel intuitive. It supports plugins, custom themes, configurable shortcuts (vim mode included), and auto-updates.

Tech: Go, Bubble Tea (TUI framework), Make + shell scripts for builds, reproducible builds via Nix flakes, distributed through Homebrew/Winget/Scoop.

amnezia-client — amnezia-vpn/amnezia-client

🔗 github.com/amnezia-vpn/amnezia-client

What it does: An open-source VPN client that one-click-deploys a personal VPN server onto your own box, then connects to it through a friendly app.

Why it matters: You want a self-hosted VPN but dread the config. Just supply server credentials and the client installs the VPN container over SSH for you. It handles OpenVPN, WireGuard, and IKEv2, adds traffic obfuscation (Cloak, Shadowsocks, AmneziaWG, XRay) and split tunneling, and covers desktop and mobile.

Tech: C++ / Qt 6.10+; OpenSSL and LibSsh; WireGuard/OpenVPN protocol implementations; Xray-core routing; Conan + CMake for builds; native Android/iOS SDKs on mobile.

---

🧭 This Week's Themes

A few patterns jumped out once everything was on one page.

AI agent harnesses are the story of the week. Five of the seventeen repos — ECC, openwork, reverse-skill, last30days-skill, and jcode — are all about giving AI coding agents structure: reusable skills, deterministic workflows, cross-session memory, routing. The frontier has clearly shifted from "can the model write code" to "can we give the model a real engineering process." That's a meaningful maturity signal for the whole agent space.

Local-first and self-hosted keeps winning hearts. AIRI (self-hosted companion), speech-to-speech (local voice), GeoLibre (local GIS), bitchat (decentralized chat), amnezia-client (self-hosted VPN), and MediaCrawler all share the same instinct — keep the data on your machine, own your stack, don't rent your privacy. It's a steady, load-bearing theme rather than a fad.

Voice and multimodal companions are maturing. AIRI and speech-to-speech both push full voice pipelines with swappable, open components. The days of voice AI being a proprietary black box are visibly ending.

Learning and curated resources still trend hard. AI-For-Beginners and awesome-systematic-trading remind us that a well-organized reading list or a solid free course can out-trend shiny new tools — sometimes people just want a map, not another territory.

---

💡 Takeaway

If you only remember one thing from this week: the agent-tooling layer is where the energy is. The winners aren't new foundation models — they're the harnesses, skills, and workflows that make existing models behave like disciplined engineers. AIRI holding the board for four straight days and ECC for three tells you developers are hungry for structure around AI, not just raw capability.

Underneath that, the local-first and self-hosted movement keeps quietly compounding. Between owning your AI companion, your voice agent, your GIS, and your VPN, the theme is consistent: people want the power and the keys. Worth keeping an eye on where these two currents meet.

See you next week. 👋

---

Compiled by Tommy Zhang | August 02, 2026