GitHub Open Source Weekly 2026-03-04: WiFi Sees Through Walls, Skills Ecosystem Explodes, Apple Neural Engine Cracked for Training

Data period: 2026-02-25 – 2026-03-04 (rolling 7 days) Sources: GitHub Trending weekly + monthly, GitHub Search API, HN Algolia

TL;DR: This week's star-gain champion is ruvnet/RuView, which uses WiFi signals for human pose estimation and added 17,166 stars in a single week. The Skills ecosystem exploded—superpowers, huggingface/skills, Agent-Skills, and learn-claude-code all charted simultaneously. The biggest New Repos highlight is maderix/ANE, which reverse-engineered Apple's private Neural Engine APIs to enable ML training with 80× the energy efficiency of an A100 GPU.

📈 Fastest Growing — Weekly Star Gains Top 10

Source: github.com/trending?since=weekly 🔁 = Also on the monthly trending list (sustained momentum signal)

#	Repo	+Stars/week	Total Stars	Language	Created
1 🔁	ruvnet/RuView	+17,166	25,799	Rust	2025-06
2 🔁	obra/superpowers	+9,010	69,641	Shell	2025-10
3	abhigyanpatwari/GitNexus	+6,522	9,312	TypeScript	2025-08
4	moeru-ai/airi	+4,365	22,389	TypeScript	2024-12
5	muratcankoylan/Agent-Skills-for-Context-Engineering	+3,826	13,231	Python	2025-12
6	alibaba/OpenSandbox	+3,800	5,517	Python	2025-12
7 🔁	bytedance/deer-flow	+3,694	24,162	Python	2025-05
8 🔁	huggingface/skills	+3,460	8,054	Python	2025-11
9	shareAI-lab/learn-claude-code	+3,122	20,962	TypeScript	2025-06
10 🔁	clockworklabs/SpacetimeDB	+2,873	22,139	Rust	2023-06

🆕 Top New Repos — This Week's Newcomers Top 10

Source: GitHub Search API (created:2026-02-25..2026-03-04, sorted by total stars)

#	Repo	Total Stars	Language	Created
1	maderix/ANE	4,741	Objective-C	2026-02-28
2	Lakr233/vphone-cli	2,931	Python	2026-02-26
3	ringhyacinth/Star-Office-UI	1,829	HTML	2026-02-26
4	smartcmd/MinecraftConsoles	1,824	C++	2026-03-01
5	ylytdeng/wechat-decrypt	1,430	Python	2026-02-28
6	ForLoopCodes/contextplus	1,307	TypeScript	2026-02-27
7	World-Open-Graph/br-acc	1,198	Python	2026-03-01
8	lochie/web-haptics	1,072	TypeScript	2026-02-25
9	open-pencil/open-pencil	1,057	TypeScript	2026-02-27
10	apple/python-apple-fm-sdk	820	Python	2026-02-25

Spotlight — Fastest Growing Top 10

📈 #1 🔁 — ruvnet/RuView｜WiFi Signals Become a Human Pose Sensor

WiFi DensePose turns commodity WiFi signals into real-time human pose estimation, vital sign monitoring, and presence detection — all without a single pixel of video.

+17,166 ★ this week｜25,799 total｜Rust｜MIT

This week's star-gain champion. RuView is based on the CMU DensePose From WiFi paper and works by analyzing perturbations in WiFi Channel State Information (CSI) to reconstruct 17 COCO keypoints and DensePose UV coordinates for human bodies. The result: through-wall sensing, breathing detection, and heartbeat monitoring—all without a camera.

Technical highlights: pure Rust implementation (~7,832 lines, zero external ML dependencies) with a MERIDIAN domain generalization system that reportedly reduces cross-environment accuracy loss from 40–70% down to under 15%, with just 10 seconds of automatic calibration. Hardware requirement: CSI-capable devices (ESP32-S3 or research-grade NICs)—standard home routers won't work.

Author ruvnet (Reuven Cohen) is a prolific open-source developer who also maintains ruvector (a vector graph neural network database, charting at #12 this week) and Ruflo (a Claude agent orchestration platform). The monthly trending appearance (🔁) confirms this isn't a passing hype spike.

What you can do with it: if you're building smart home, elder care monitoring, or search-and-rescue systems, this repo is worth a deep dive. Note the hardware barrier—this isn't "any off-the-shelf router will do."

📈 #2 🔁 — obra/superpowers｜Two Months Dominating the Skills Framework Chart

An agentic skills framework & software development methodology that works.

+9,010 ★ this week｜69,641 total｜Shell｜MIT

Jesse Vincent's superpowers has been on GitHub Trending for two consecutive months, adding another 9,010 stars this week and approaching 70,000 total. The core isn't a new AI model—it's a structural constraint system: markdown-formatted Skill files that force AI coding agents (Claude Code, Cursor, Codex) to follow a workflow of "clarify requirements first → generate a spec → write tests before code → then implement."

Since becoming an official Anthropic plugin on January 15, 2026, adoption has accelerated. Three months of growth totaling 27,000 stars translates to a steady ~9,000 stars per month. A monthly trending regular (🔁)—this isn't hype, it's genuine production usage.

📈 #3 — abhigyanpatwari/GitNexus｜A Code Knowledge Graph That Runs in Your Browser

The Zero-Server Code Intelligence Engine — client-side knowledge graph creator with built-in Graph RAG Agent.

+6,522 ★ this week｜9,312 total｜TypeScript｜Created 2025-08

GitNexus runs entirely in the browser with no server required. Drop in a GitHub repo or ZIP file and it builds a local code knowledge graph using KuzuDB + Tree-sitter, then lets you explore the codebase structure in natural language through a Graph RAG Agent.

The key differentiator: rather than raw RAG retrieval at query time, it pre-computes function clusters and execution paths during indexing. Combined with BM25 + semantic + RRF search, the AI agent gets high-fidelity context in a single pass. Supports MCP server integration with Cursor and Claude Code.

+6,522 stars in a week is a remarkable number, especially for an individual project without a major company behind it.

📈 #4 — moeru-ai/airi｜Self-Hosted AI Companion That Plays Minecraft

Self hosted, you-owned Grok Companion — capable of realtime voice chat, Minecraft, Factorio playing.

+4,365 ★ this week｜22,389 total｜TypeScript｜MIT

airi is a self-hosted AI companion framework aimed at letting you own your own digital life. It supports real-time voice chat, Minecraft and Factorio gameplay, Live2D / VRM avatars, and runs across Web, macOS, and Windows.

Its positioning is distinctive: it explicitly targets the same space as Neuro-sama (the well-known AI VTuber), while emphasizing "you own your AI companion." With 22,389 stars, there's clearly a committed developer community pushing the AI companion category forward.

📈 #5 — muratcankoylan/Agent-Skills-for-Context-Engineering｜A Skill Library Purpose-Built for Context Engineering

A comprehensive collection of Agent Skills for context engineering, multi-agent architectures, and production agent systems.

+3,826 ★ this week｜13,231 total｜Python｜MIT

Charting alongside obra/superpowers and huggingface/skills, this forms this week's Skills triangle. The focus is "Context Engineering"—how to effectively manage context windows in multi-agent systems: compression strategies, context routing, and debugging methodologies.

If you're building complex AI agent systems and context management is your bottleneck, this is the most technically relevant repo on the chart this week.

📈 #6 — alibaba/OpenSandbox｜Alibaba's Open-Source AI Agent Sandbox Platform

General-purpose sandbox platform for AI applications — Docker/Kubernetes runtimes for Coding Agents, GUI Agents, Agent Evaluation, AI Code Execution, and RL Training.

+3,800 ★ this week｜5,517 total｜Python｜Apache-2.0

Alibaba's open-source general-purpose AI agent sandbox. Unlike E2B (cloud-first) or Modal (serverless functions), OpenSandbox offers full deployment flexibility from local Docker to Kubernetes clusters, with Python, TypeScript, and Java SDKs—with particular emphasis on GUI agent scenarios (VNC remote desktop environments).

Already integrated with LangGraph, Claude Code, Gemini CLI, and Playwright. For agent developers who need secure isolated execution environments, this is the most complete open-source option available. MarkTechPost has detailed coverage.

📈 #7 🔁 — bytedance/deer-flow｜ByteDance's Long-Running Task Agent Framework

An open-source SuperAgent harness that researches, codes, and creates — handles tasks from minutes to hours.

+3,694 ★ this week｜24,162 total｜Python｜MIT

DeerFlow is ByteDance's open-source SuperAgent framework built on LangGraph + LangChain, designed to solve a problem most agent frameworks ignore: autonomous tasks that run for hours. Most agent frameworks assume second-scale request-response interactions. DeerFlow includes built-in sandboxed execution, persistent memory, and sub-agent spawning to handle deep research and large-scale refactoring that takes hours.

A monthly trending regular (🔁) with 24k stars. If your agent tasks run longer than 10 minutes, DeerFlow is worth evaluating.

📈 #8 🔁 — huggingface/skills｜HuggingFace's Official Skills Registry

HuggingFace's official cross-platform Agent Skills repository

+3,460 ★ this week｜8,054 total｜Python｜Apache-2.0

HuggingFace's official skills repository, providing standardized skill definitions for ML workflows: model fine-tuning, DPO alignment, GGUF conversion, and dataset creation. Supports Claude Code, Codex, Gemini CLI, Cursor and other platforms.

How it differs from obra/superpowers: superpowers focuses on software development discipline (TDD, requirements clarification); huggingface/skills focuses on ML training workflows. The two are complementary. A monthly trending regular (🔁)—the Skills ecosystem is officially moving from individual tool to platform standard.

📈 #9 — shareAI-lab/learn-claude-code｜Build a Claude Code-Like Agent from Scratch

Bash is all you need — A nano Claude Code–like agent, built from 0 to 1

+3,122 ★ this week｜20,962 total｜TypeScript｜MIT

A learning-oriented project that walks you through understanding from scratch how AI coding agents like Claude Code actually work. The core philosophy is "Bash is all you need"—using a minimal implementation to help learners grasp the fundamentals of agent architecture.

With 20,962 stars and 4,129 forks, the demand for understanding how AI agents work internally is clearly strong. This week's +3,122 stars likely got a boost from the overall Skills ecosystem momentum.

📈 #10 🔁 — clockworklabs/SpacetimeDB｜Database 2.0 for Games, Benchmark Controversy Follows

Development at the speed of light

+2,873 ★ this week｜22,139 total｜Rust｜Created 2023-06

SpacetimeDB is a real-time backend framework that merges database and server logic into a single layer. You write "modules" in TypeScript, C#, or Rust directly inside the database; clients connect to the database directly; all state lives in memory with WAL persistence. Its production case study is the full backend of the MMORPG BitCraft Online.

After the 2.0 release, the benchmark results sparked HN debate: the team claimed 170k TPS (Rust module), and the community questioned whether comparisons were fair (different clients and protocols). The team subsequently published a fair benchmark showing 2.3× PostgreSQL under equivalent conditions and 8.3× at high load. The community responded positively to this transparency.

A monthly trending regular (🔁)—if you're building real-time multi-user applications (games, collaborative tools), SpacetimeDB is a non-conventional option worth evaluating.

Spotlight — Top New Repos Top 10

🆕 #1 — maderix/ANE｜Reverse-Engineering Apple Neural Engine to Enable ML Training

Training neural networks on Apple Neural Engine via reverse-engineered private APIs

4,741 total ★｜Objective-C｜MIT｜Created 2026-02-28

The most technically deep project in this week's New Repos. The author reverse-engineered Apple's private _ANEClient / _ANECompiler APIs and the MIL (Model Intermediate Language) format to enable ML training on the Apple Neural Engine for the first time. Apple has only ever exposed the ANE for inference—training support has never been officially offered.

According to the author's measurements, the M4's ANE delivers 19 TFLOPS FP16 at 2.8W—an energy efficiency of 6.6 TFLOPS/W, or 80× better than an A100 GPU. A 109M parameter Llama2 transformer was successfully trained on ANE. The legal basis cites Sega v. Accolade (1992) and DMCA §1201(f) interoperability provisions; no Apple proprietary code was used. Apple has not responded as of publication.

What this means for Mac users: if Apple eventually decides to officially open ANE training APIs, it would fundamentally change the on-device ML landscape.

🆕 #2 — Lakr233/vphone-cli｜Full iOS VM on Apple Silicon

2,931 total ★｜Python｜MIT｜Created 2026-02-26

A command-line tool for running a full iOS 26 virtual machine (not a simulator) on Apple Silicon Macs. Paired with 34306/vphone-aio (a pre-built 12GB VM image for one-click launch) and wh1te4ever/super-tart-vphone-writeup (technical documentation), three vphone-related repos charted simultaneously this week.

Why it matters: iOS security research and QA testing have long depended on paid Corellium licenses (running thousands of dollars per year). vphone provides a local, free alternative.

🆕 #3 — ringhyacinth/Star-Office-UI｜A Pixel-Art Office Dashboard for OpenClaw

A pixel office for your OpenClaw — turn invisible work states into a cozy little space with characters, daily notes, and guest agents.

1,829 total ★｜HTML｜Created 2026-02-26

Visualizes an OpenClaw agent's work state as a pixel-art office interface built with the Phaser game engine. Features include characters, daily notes, and guest agents. Art assets are limited to non-commercial educational use.

The technical depth here isn't the point—what this repo represents is the OpenClaw ecosystem moving from "tool" toward "experience."

🆕 #4 — smartcmd/MinecraftConsoles｜A C++ Minecraft Reimplementation

A certain block game

1,824 total ★｜C++｜Created 2026-03-01

The description is deliberately understated: "a certain block game." A C++ reimplementation of Minecraft that pulled 1,824 stars and 350 forks in three days. Worth a look for anyone interested in game engine internals and C++ implementation techniques.

🆕 #5 — ylytdeng/wechat-decrypt｜WeChat 4.0 Database Decryptor

WeChat 4.0 database decryptor — extract keys from memory, decrypt SQLCipher 4 databases, real-time message monitor

1,430 total ★｜Python｜Created 2026-02-28

Extracts encryption keys from WeChat 4.0, decrypts SQLCipher 4 databases, and supports real-time message monitoring. The 1,087 forks (75% of star count) signal extremely high intent to use.

Note: Tools of this type may implicate privacy regulations in various jurisdictions. Verify compliance before use.

🆕 #6 — ForLoopCodes/contextplus｜MCP Server That Helps AI Truly Understand Large Codebases

Context+ is an MCP server combining RAG, Tree-sitter AST, Spectral Clustering, and Obsidian-style linking to turn massive codebases into searchable feature graphs.

1,307 total ★｜TypeScript｜MIT｜Created 2026-02-27

Context+ is an MCP server that converts large codebases into a searchable hierarchical feature graph through RAG + Tree-sitter AST analysis + spectral clustering + Obsidian-style bidirectional linking.

Together with GitNexus (#3), this forms this week's dual-track "code understanding" theme: GitNexus is a browser-side visual knowledge graph, Context+ is an MCP server-side semantic index. Both are solving the same problem: giving AI agents a more complete understanding of large codebases.

🆕 #7 — World-Open-Graph/br-acc｜World Transparency Graph

1,198 total ★｜Python｜AGPL-3.0｜Created 2026-03-01

The public codebase for World Transparency Graph, a site still in development. With limited description and no verifiable use-case documentation at this time, the best approach is to watch the repo directly for updates.

🆕 #8 — lochie/web-haptics｜Haptic Feedback for the Mobile Web

Haptic feedback for the mobile web

1,072 total ★｜TypeScript｜MIT｜Created 2026-02-25

A lightweight library that adds haptic feedback (vibration) to mobile websites. Supports React, Vue, Svelte, and vanilla JS with a simple API (trigger("success")). The live demo lets you try it directly on a phone.

What this means for frontend developers: mobile web interactions have long lagged behind native apps. web-haptics closes that gap with the smallest possible API surface.

🆕 #9 — open-pencil/open-pencil｜AI-Native Open-Source Figma Alternative

AI-native design editor. Open-source Figma alternative.

1,057 total ★｜TypeScript｜MIT｜Created 2026-02-27

Open Pencil is a fully local design editor with bidirectional .fig file compatibility (fills, strokes, auto-layout, effects, vector networks) and a built-in AI assistant (via OpenRouter, no backend required). Approximately 14.5K lines of TypeScript, rendered with CanvasKit.

Still early-stage and far from Figma's feature completeness. But the combination of "AI-native + local-first + .fig compatible" is a meaningful differentiation angle.

🆕 #10 — apple/python-apple-fm-sdk｜Apple's Official Python SDK for On-Device Foundation Models

Python bindings for access to the on-device model at the core of Apple Intelligence through the Foundation Models framework

820 total ★｜Python｜Apache-2.0｜Created 2026-02-25

Apple's officially released Python SDK (Beta, requires Xcode 26.0+) that lets developers access on-device Foundation Models via SystemLanguageModel() and LanguageModelSession(). Supports text generation and function calling, meaning you can build local AI agents.

An interesting contrast with maderix/ANE (#1): one is a community reverse-engineering effort, the other is an official Apple release. Both point in the same direction: the on-device AI developer toolchain is maturing rapidly.

Monthly Trend Comparison

This week had 5 repos on the monthly trending list simultaneously (🔁)—the most in recent weeks:

Repo	Weekly gain	Monthly signal
ruvnet/RuView	+17,166	Monthly #4 (+18,864)
obra/superpowers	+9,010	Monthly #9 (+26,060), third consecutive month
bytedance/deer-flow	+3,694	Monthly #10 (+4,299)
huggingface/skills	+3,460	Monthly #8 (+6,885)
clockworklabs/SpacetimeDB	+2,873	Monthly #14 (+3,129)

The monthly chart's top two are openclaw/openclaw (+99,675) and KeygraphHQ/shannon (+22,627, an AI automated vulnerability hunting hacker), signaling that the AI agent ecosystem hype cycle is still in full expansion.

Weekly Trend Insights

The Skills ecosystem has reached full explosion—from frameworks to curriculum

This week's Fastest Growing Top 10 includes four repos directly related to Skills: obra/superpowers (framework), huggingface/skills (platform registry), Agent-Skills-for-Context-Engineering (vertical skills), and learn-claude-code (education). Add the monthly chart entries—openai/skills (Codex Skills), clawhub (OpenClaw Skill Directory), and gsd-build/get-shit-done (spec-driven development)—and Skills has moved from late-2025 proof-of-concept to full ecosystem in 2026. The next question: who builds the npm for Skills and becomes the central registry and distribution mechanism?

Edge AI and on-device ML: a breakout year

RuView (WiFi sensing, Rust), Moonshine (edge ASR, HN 316 pts, more accurate than Whisper Large V3 and 5–15× faster), maderix/ANE (Apple Neural Engine training), and apple/python-apple-fm-sdk (official on-device SDK) all appeared the same week, each pushing AI from the cloud toward the device from a different angle. The simultaneous appearance of ANE's reverse engineering and Apple's official SDK is especially symbolic—the on-device ML developer toolchain is in a period of rapid maturation.

Major tech companies are open-sourcing their agent infrastructure

alibaba/OpenSandbox (sandboxing) and bytedance/deer-flow (long-running task agents) charted in the same week, alongside monthly chart entry github/gh-aw (GitHub Agentic Workflows). Big companies are now open-sourcing the agent infrastructure they've built internally. For independent developers, this is good news: building production-grade agent systems no longer requires starting from scratch—you can stand directly on the shoulders of what the big players have already built.