GitHub Open Source Weekly 2026-03-11: Karpathy's Return Sparks Research Automation, Skills Ecosystem Blooms, OSINT Tool Hits 304 HN Points
Data Period: 2026-03-04 to 2026-03-11 (Rolling 7 days) Sources: GitHub Trending weekly + monthly, GitHub Search API, HN Algolia, Twitter/X
TL;DR: The biggest surprise this week is karpathy/autoresearch, reaching 22,983 stars in three days with HN's highest discussion score at 198 points. The Skills ecosystem saw explosive growth with 5 of the Top 10 New Repos directly related to Skills. The biggest non-AI dark horse is OSINT tool Shadowbroker, integrating 15 real-time data sources from corporate jet tracking to satellite orbits, claiming the week's HN discussion crown at 304 points.
📈 Fastest Growing — Weekly Star Gains Top 14
Source:
github.com/trending?since=weekly🔁 = Also appears on monthly trending (sustained momentum signal)
| # | Project | +Stars/Week | Total Stars | Language | Created |
|---|---|---|---|---|---|
| 1 | msitarzewski/agency-agents | +19,856 | 25,639 | Shell | 2025-10 |
| 2 🔁 | moeru-ai/airi | +10,249 | 32,188 | TypeScript | 2024-12 |
| 3 🔁 | ruvnet/RuView | +9,844 | 34,353 | Rust | 2025-06 |
| 4 | 666ghj/MiroFish | +8,983 | 14,321 | Python | 2025-11 |
| 5 🔁 | bytedance/deer-flow | +4,339 | 28,617 | Python | 2025-05 |
| 6 🔁 | shareAI-lab/learn-claude-code | +4,137 | 24,909 | TypeScript | 2025-06 |
| 7 | ItzCrazyKns/Vane | +3,187 | 32,643 | TypeScript | 2024-04 |
| 8 | GoogleCloudPlatform/generative-ai | +2,810 | 15,771 | Jupyter | 2023-05 |
| 9 🔁 | alibaba/OpenSandbox | +2,349 | 7,407 | Python | 2025-12 |
| 10 | teng-lin/notebooklm-py | +2,068 | 4,735 | Python | 2026-01 |
| 11 | QwenLM/Qwen-Agent | +1,835 | 15,345 | Python | 2023-09 |
| 12 | NousResearch/hermes-agent | +1,814 | 3,831 | Python | 2025-07 |
| 13 | alirezarezvani/claude-skills | +1,561 | 4,014 | Python | 2025-10 |
| 14 | inclusionAI/AReaL | +1,018 | 4,634 | Python | 2025-02 |
🆕 Top New Repos — Born This Week Top 10
Source: GitHub Search API (
created:2026-03-04..2026-03-11, sorted by total stars)
| # | Project | Total Stars | Language | Created |
|---|---|---|---|---|
| 1 | karpathy/autoresearch | 22,983 | Python | 2026-03-06 |
| 2 | HKUDS/CLI-Anything | 2,707 | Python | 2026-03-08 |
| 3 | twostraws/SwiftUI-Agent-Skill | 1,754 | — | 2026-03-05 |
| 4 | duoan/TorchCode | 1,520 | Jupyter | 2026-03-04 |
| 5 | jackwener/twitter-cli | 1,463 | Python | 2026-03-05 |
| 6 | BigBodyCobain/Shadowbroker | 1,415 | TypeScript | 2026-03-05 |
| 7 | viperrcrypto/Siftly | 1,401 | TypeScript | 2026-03-04 |
| 8 | cyxzdev/Uncodixfy | 1,357 | — | 2026-03-06 |
| 9 | ParthJadhav/app-store-screenshots | 1,017 | — | 2026-03-07 |
| 10 | FreedomIntelligence/OpenClaw-Medical-Skills | 927 | Python | 2026-03-08 |
Spotlight — Fastest Growing Top 14
📈 #1 — msitarzewski/agency-agents | A Complete AI Agency: From Frontend Engineers to Community Ninjas
A complete AI agency at your fingertips - From frontend wizards to Reddit community ninjas, from whimsy injectors to reality checkers. Each agent is a specialized expert with personality, processes, and proven deliverables.
This Week +19,856 ★ | Total ★25,639 | Shell | MIT
agency-agents provides a set of pre-defined AI Agent roles covering frontend engineering, marketing, social media, fact-checking, and more. Each agent comes with its own personality, workflow, and deliverable format. Technically driven by Shell scripts, it can connect to various LLM backends.
This week's gain of nearly 20,000 stars stands in sharp contrast to just 2 HN points. The Twitter story was the complete opposite: @RoundtableSpace's tweet "Someone built a complete AI agency on GitHub with 51 specialized agents" pulled 11,045 likes, making it one of the week's most viral open source tweets. However, @OneManSaas pointed out: "10K stars is interesting, but these complete setups always skip the hardest part: actually making money."
This gap illustrates the disconnect between social media virality and technical community endorsement. Worth a look if you need to quickly assemble multi-role agent workflows, but don't expect deep technical substance.
📈 #2 🔁 — moeru-ai/airi | Self-Hosted AI VTuber Companion That Plays Minecraft
Self hosted, you-owned Grok Companion, a container of souls of waifu, cyber livings to bring them into our worlds, wishing to achieve Neuro-sama's altitude.
This Week +10,249 ★ | Total ★32,188 | TypeScript | MIT
airi aims to replicate the VTuber Neuro-sama experience: an AI character with personality that can voice chat in real-time, play Minecraft and Factorio, deployable on Web, macOS, and Windows. It uses Live2D and VRM models with Grok API as the conversation backend, emphasizing "self-hosted, you own everything."
This is its second consecutive week on the chart (+10,249, also on monthly trending), indicating sustained growth momentum. The AI VTuber/companion space still has explosive potential, especially the privacy-first, self-hosted direction that's building a loyal user base.
📈 #3 🔁 — ruvnet/RuView | WiFi See-Through-Walls Controversy: Community Says "Nobody Can Run It"
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.
This Week +9,844 ★ | Total ★34,353 | Rust | MIT
RuView claims to perform human pose estimation and vital sign monitoring using ordinary WiFi router signals, no cameras needed. The technical basis is CMU's 2023 DensePose From WiFi paper. Last week's growth leader (+17,166), it holds third place this week at +9,844.
Notably, HN saw an explicitly critical post: Top trending repo claims to detect movement via WiFi, yet no one can run it (9 points). Twitter reactions were polarized: @The_Cyber_News (336 likes) and @heygurisingh (159 likes) went viral with "see through walls using WiFi" headlines, sparking privacy concerns about "zero-cost surveillance." Even @grok (573 likes) weighed in to explain the technical principles.
Community concerns include: incomplete documentation, questionable demo videos, and the gap between the repo and the academic paper. If you're thinking about deploying this, check the Issues and HN discussion first. It's still more concept than usable tool.
📈 #4 — 666ghj/MiroFish | Swarm Intelligence Prediction Engine
A Simple and Universal Swarm Intelligence Engine, Predicting Anything.
This Week +8,983 ★ | Total ★14,321 | Python | AGPL-3.0
MiroFish is a swarm intelligence framework using multi-agent simulation for predictions, covering financial forecasting, public sentiment analysis, and crowd behavior estimation. It combines knowledge graphs with LLMs to simulate collective responses to events.
The AGPL-3.0 license means commercial use requires open-sourcing your code, and with no commits in nearly two months (last push 2026-03-07), the maintenance status needs watching. Good for researchers exploring swarm intelligence and multi-agent simulation, but evaluate maintenance risk before production deployment.
📈 #5 🔁 — bytedance/deer-flow | ByteDance's SuperAgent, Three Weeks of Sustained Momentum
An open-source SuperAgent harness that researches, codes, and creates. With the help of sandboxes, memories, tools, skills and subagents, it handles different levels of tasks.
This Week +4,339 ★ | Total ★28,617 | Python | MIT
DeerFlow is ByteDance's open-source SuperAgent framework supporting sandbox execution, memory, tool calling, and sub-agent collaboration, designed for complex tasks requiring minutes to hours. Three consecutive weeks on the monthly trending chart shows ByteDance's sustained promotion efforts combined with broad feature coverage (research, coding, content creation) that attracts users.
On Twitter, the official LangChain account directly endorsed DeerFlow (334 likes), and @jasonzhou1993 listed it as the top pick in their "most noteworthy AI Agent projects this week" thread (1,002 likes). For developers building long-running agent workflows, DeerFlow is among the most feature-complete open-source options available.
📈 #6 🔁 — shareAI-lab/learn-claude-code | Understanding Claude Code's Internals Through Bash
Bash is all you need - A nano Claude Code–like agent, built from 0 to 1
This Week +4,137 ★ | Total ★24,909 | TypeScript | MIT
learn-claude-code is an educational repo that builds a nano version of Claude Code from scratch using Bash, helping developers understand the underlying principles of AI Coding Agents. For those already using Claude Code but wanting to know "what's actually happening under the hood," this is the most direct deconstruction available. Four consecutive weeks on the chart with sustained monthly trending shows clear educational demand.
📈 #7 — ItzCrazyKns/Vane | AI-Powered Search Engine
Vane is an AI-powered answering engine.
This Week +3,187 ★ | Total ★32,643 | TypeScript | MIT
Vane is an open-source AI search engine (formerly Perplexica), integrating SearxNG for meta-search with a RAG architecture supporting multiple LLM backends, fully self-hostable. This week's +3,187 growth came primarily from GitHub community sharing rather than HN (all three HN hits were keyword collisions). For those wanting a self-hosted Perplexity alternative, Vane is one of the most mature options available.
📈 #8 — GoogleCloudPlatform/generative-ai | Google's Official Gemini Sample Collection Gets Major Update
Sample code and notebooks for Generative AI on Google Cloud, with Gemini on Vertex AI
This Week +2,810 ★ | Total ★15,771 | Jupyter Notebook | Apache-2.0
GoogleCloudPlatform/generative-ai is Google Cloud's official Gemini + Vertex AI sample repo, boosted this week by Google's launch of the Always-On Memory Agent feature (HN 5 points). The most authoritative sample source for developers building AI applications on GCP.
📈 #9 🔁 — alibaba/OpenSandbox | Alibaba's AI Agent Sandbox, Sustained Monthly Momentum
General-purpose sandbox platform for AI applications, offering multi-language SDKs, unified sandbox APIs, Docker/Kubernetes runtimes.
This Week +2,349 ★ | Total ★7,407 | Python | Apache-2.0
OpenSandbox provides unified sandbox APIs letting AI agents safely execute code, operate GUIs, and run RL training, with Docker and Kubernetes deployment support. From Alibaba, with sustained monthly trending momentum, it's a representative example of big tech's open-source AI infrastructure push.
📈 #10 — teng-lin/notebooklm-py | Full NotebookLM Control via Python and Claude Skill
Unofficial Python API and agentic skill for Google NotebookLM. Full programmatic access to NotebookLM's features.
This Week +2,068 ★ | Total ★4,735 | Python | MIT
notebooklm-py provides an unofficial Python API for NotebookLM, including CLI and Claude Skill interfaces, accessing all NotebookLM features including hidden capabilities not available in the Web UI. Currently the only publicly available automation solution for integrating NotebookLM into agent workflows.
📈 #11 — QwenLM/Qwen-Agent | Alibaba's Qwen 3.0 Agent Framework with MCP Support
Agent framework and applications built upon Qwen >= 3.0, featuring Function Calling, MCP, Code Interpreter, RAG, Chrome extension.
This Week +1,835 ★ | Total ★15,345 | Python | Apache-2.0
Qwen-Agent is the official agent framework for Qwen 3.0, boosted this week by the Qwen 3.0 release. Supports MCP protocol, Function Calling, Code Interpreter, RAG, and has a Chrome extension. The 455 open issues is a relatively high number; evaluate known issues before adoption.
📈 #12 — NousResearch/hermes-agent | The Personal AI Agent "That Grows With You"
The agent that grows with you
This Week +1,814 ★ | Total ★3,831 | Python | MIT
Hermes Agent is NousResearch's personal AI agent emphasizing accumulated personalized memory over time, with a design philosophy of long-term companionship rather than one-off tasks. NousResearch is known for their Hermes series fine-tuned models; this is their first complete agent product. HN had two threads (3 points + 1 point), with the community taking a wait-and-see approach.
📈 #13 — alirezarezvani/claude-skills | 180+ Production-Ready Claude Code Skills Marketplace
+180 production-ready skills & plugins for Claude Code, OpenAI Codex, and OpenClaw.
This Week +1,561 ★ | Total ★4,014 | Python | MIT
claude-skills curates over 180 ready-to-use Skills covering engineering, marketing, product, compliance, and C-level advisory roles, installable via /plugin marketplace. The HN Show HN post earned 13 points and 20 comments, with some questioning quality consistency ("How do you ensure all 180 Skills are good?") while others praised the coverage. Twitter's Skills ecosystem discussion was also active: @PawelHuryn (2,235 likes) shared PM-related Skills usage experiences, while @Baconbrix (Expo official, 1,953 likes) demonstrated integrating Skills into React Native development workflows.
📈 #14 — inclusionAI/AReaL | Lightning-Fast RL Framework for LLMs
Lightning-Fast RL for LLM Reasoning and Agents. Made Simple & Flexible.
This Week +1,018 ★ | Total ★4,634 | Python | Apache-2.0
AReaL provides a reinforcement learning framework optimized for LLM reasoning and agent training, emphasizing speed and flexibility. Suited for AI research teams building their own RLHF or RLAIF pipelines. InclusionAI is a Chinese AI company with an active open-source stance.
Spotlight — Top New Repos Top 10
🆕 #1 — karpathy/autoresearch | Karpathy's New Project: Let AI Research on a Single GPU
AI agents running research on single-GPU nanochat training automatically
★22,983 | Python | Created 2026-03-06
This is the week's most noteworthy new repo. Andrej Karpathy (former Tesla AI Director, OpenAI co-founder) released autoresearch on March 6, enabling AI agents to automatically run nanochat (small chat model) training research on a single GPU. It reached 22,983 stars in three days.
HN discussion hit 198 points with 56 comments, the highest engagement among all repos this week. Core discussion topics include: whether this approach can truly automate ML research loops, whether single-GPU is too limiting, and comparisons with prior work like AI Scientist. The community broadly agrees that Karpathy's reputation is itself an endorsement; despite the early stage, community members have already ported it to Windows/RTX GPU versions.
Twitter amplification was even more dramatic: Karpathy's own launch tweet pulled 28,072 likes, with a follow-up results tweet getting 19,006 likes. @Chris_Worsey approached it from a market application angle (3,913 likes), noting that the significance for small research teams is "doing with one consumer GPU what used to require an entire lab."
For ML engineers interested in research automation, autoresearch is currently the most representative "one person, one GPU, let AI do the research" experiment.
🆕 #2 — HKUDS/CLI-Anything | Making All Software Agent-Native
CLI-Anything: Making ALL Software Agent-Native
★2,707 | Python | Created 2026-03-08
CLI-Anything comes from the University of Hong Kong (HKUDS), aiming to wrap any CLI-enabled software into tools AI agents can directly invoke. Author @huang_chao4969 shared the design philosophy on Twitter (860 likes), and @alifcoder sparked 636 likes with their "Making all software Agent-Native" post. This problem (how to let agents use legacy tools) is precisely one of the MCP ecosystem's core challenges, making this academically-backed tool solution worth tracking.
🆕 #3 — twostraws/SwiftUI-Agent-Skill | Paul Hudson Builds a SwiftUI Skill for Claude Code
SwiftUI agent skill for Claude Code, Codex, and other AI tools.
★1,754 | MIT | Created 2026-03-05
Author Paul Hudson is the founder of Hacking with Swift, the world's most popular Swift tutorial platform reaching over a million developers annually. SwiftUI-Agent-Skill provides a SwiftUI development Skill for Claude Code and Codex, equipping AI agents with iOS development domain knowledge. This is the first contribution from a top-tier community educator to the Skills ecosystem, signaling that the Skills format has entered the "knowledge author" market.
🆕 #4 — duoan/TorchCode | LeetCode for PyTorch: Build GPT-2 and Attention From Scratch
LeetCode for PyTorch — practice implementing softmax, attention, GPT-2 and more from scratch with instant auto-grading.
★1,520 | Jupyter Notebook | Created 2026-03-04
TorchCode is an interactive platform for practicing low-level PyTorch implementations, with exercises covering softmax, attention, GPT-2, and other core components. Jupyter-based with self-hosting support and instant auto-grading. For engineers wanting to deeply understand Transformer internals, prep for ML interviews, or just "hand-code GPT once," this is a rare tool that's both fun and practical.
🆕 #5 — jackwener/twitter-cli | Browse Twitter/X From Your Terminal
A CLI for Twitter/X — feed, bookmarks, and user timeline in terminal
★1,463 | Python | Created 2026-03-05
twitter-cli lets you view Twitter/X feeds, bookmarks, and user timelines in the terminal. Well-known cryptography expert @jedisct1 (libsodium author) endorsed the tool on Twitter (1,113 likes), driving developer community attention. Steady demand for this type of tool continues, especially from privacy-conscious developers or those who prefer terminal workflows. Its appearance alongside Siftly (another Twitter bookmark tool this week) confirms real demand among developers dissatisfied with Twitter/X's official UI.
🆕 #6 — BigBodyCobain/Shadowbroker | OSINT Dashboard Integrating 15 Real-Time Sources, 304 HN Points
Open-source intelligence for the global theater. Track everything from corporate jets, spy satellites, to seismic events in one unified interface.
★1,415 | TypeScript | Created 2026-03-05
Shadowbroker is the week's hottest non-AI tool on HN, hitting 304 points with 120 comments. It integrates 15 public real-time data sources including corporate jet tracking, spy satellite orbits, and global seismic events into a unified open-source OSINT dashboard.
The core HN debate centers on "where do the boundaries of public intelligence lie?" Some see the integration itself as democratizing capabilities that previously required professional resources, while others worry about misuse risks. Twitter was equally enthusiastic: @GithubProjects pulled 1,832 likes, @tom_doerr (602 likes) called it "a game changer for open-source intelligence tools," and @RoundtableSpace (409 likes) shared project screenshots sparking discussion. A fascinating data integration tool for security researchers, journalists, and geopolitics enthusiasts.
🆕 #7 — viperrcrypto/Siftly | Local AI-Powered Twitter/X Bookmark Organizer
Local Twitter/X bookmark organizer with AI categorization and mindmap visualization
★1,401 | TypeScript | MIT | Created 2026-03-04
Siftly uses local AI to organize your Twitter/X bookmarks with automatic categorization and mindmap visualization. Local-first design keeps your data on your machine. Worth trying for heavy Twitter/X users with unmanageable bookmark chaos.
🆕 #8 — cyxzdev/Uncodixfy | Fighting the "Cursor Look" with Prompt Engineering
the holly uncodexify instructions - letting GPT create uncodexified UI
★1,357 | Created 2026-03-06
Uncodixfy is a set of system prompts that make GPT generate UIs that don't look AI-generated. It addresses a real pain point: the flood of AI-generated interfaces sharing the same template aesthetic (often called the "Cursor look" or "v0 look"). Uncodixfy tries to break this pattern. 1,357 stars confirms this pain point is widespread.
🆕 #9 — ParthJadhav/app-store-screenshots | AI-Generated App Store Screenshots
end to end app store screenshot creation using AI
★1,017 | Created 2026-03-07
app-store-screenshots provides a complete App Store screenshot generation Skill, integrating Claude, Cursor, and other AI tools to automate the pre-launch screenshot workflow for iOS apps. Screenshot creation is a universally dreaded task for iOS developers, and this tool directly addresses the pain point.
🆕 #10 — FreedomIntelligence/OpenClaw-Medical-Skills | Largest Open-Source Medical AI Skills Library
The largest open-source medical AI skills library for OpenClaw.
★927 | Python | Created 2026-03-08
OpenClaw-Medical-Skills is an open-source contribution from the Chinese University of Hong Kong's Faculty of Medicine, providing a medical AI Skills library for OpenClaw. This signals the Skills ecosystem extending into vertical domains (healthcare), with academic institutions bringing highly specialized knowledge.
Monthly Trending Comparison
Five of this week's Fastest Growing also appear on monthly trending (🔁):
| Project | Monthly Gain | Notes |
|---|---|---|
| ruvnet/RuView | +28,314 | Three-week monthly leader, controversies persist |
| moeru-ai/airi | +14,824 | VTuber AI momentum continues |
| bytedance/deer-flow | +8,378 | Strong ByteDance promotion |
| shareAI-lab/learn-claude-code | +8,053 | Claude Code education demand holds |
| alibaba/OpenSandbox | +6,316 | Sustained big tech AI infrastructure interest |
Notable monthly chart entries not in this week's weekly:
- openclaw/openclaw: Monthly +117,792, the most extreme number on the monthly chart, showing OpenClaw's ecosystem dominance over the past 30 days
- obra/superpowers: Monthly +27,391, Jesse Vincent's agentic skills framework maintains momentum
- github/gh-aw: GitHub's official Agentic Workflows tool, monthly +2,923, representing GitHub's active push into agent workflow integration
Trend Insights
Skills Ecosystem Enters Vertical Specialization Phase
Last week Skills claimed four spots. This week, five of the Top 10 New Repos directly relate to Skills (SwiftUI-Agent-Skill, app-store-screenshots, OpenClaw-Medical-Skills, Claude-to-IM-skill, and Uncodixfy's prompt tooling logic). The more significant qualitative shift: from "general-purpose Skills packages" (agency-agents, claude-skills) toward "vertical domain Skills" (SwiftUI, healthcare). The entry of top-tier domain educators like Paul Hudson signals the Skills format has earned enough credibility for knowledge authors to invest seriously.
The Karpathy Effect: Personal Brand Drives Research Direction
autoresearch hit 22,983 stars and 198 HN points in three days, directly tied to Karpathy's personal influence. Other technically similar research automation tools launched simultaneously but saw nowhere near the same traction. In the AI tools space, founder brand remains the strongest cold-start signal, meaning "who's building it" sometimes matters more than "what it is" for initial attention.
Non-AI Tools Outperform AI on HN, but Twitter Tells a Different Story
Shadowbroker (304 points) far exceeded growth leader agency-agents (2 points) in HN engagement. But Twitter painted the opposite picture: agency-agents' viral tweet hit 11,045 likes, Karpathy's autoresearch reached 28,072 likes, while Shadowbroker, despite being the HN champion, saw relatively scattered Twitter discussion. This "HN vs Twitter" gap reveals fundamental differences between the two platform communities: HN favors substantive technical depth and ethical debates (OSINT, privacy), while Twitter is the arena for brand effects and viral reach. For open-source developers, gaining attention on both platforms requires different strategies.
