Every month, a new wave of GitHub repos explodes from obscurity into tens of thousands of stars. Most of them are forgettable. This month? Not so much. The March 2026 trending list reads like a manifesto for the agentic AI era: personal AI assistants, swarm intelligence engines, WiFi-based human detection, and tools to build your own Claude Code clone from scratch.
I went through the 10 fastest-growing GitHub projects of the month, checked every repo, and here’s what’s actually worth your attention.
1. OpenClaw: The 300K-Star Personal AI Assistant
openclaw/openclaw crossed 313,000 stars this month, making it the single fastest-growing repo on the platform. Written in TypeScript, it’s a self-hosted personal AI assistant that runs 24/7 on any OS. The pitch: you own your data, you own your agent, no cloud dependency.
The “lobster way” branding is… a choice. But the project is legit. It’s become the default scaffold for people who want to run persistent AI agents on their own hardware. If you’ve been thinking about setting up a local AI assistant that doesn’t phone home to OpenAI, this is the one everyone’s picking.
2. Superpowers: Agentic Skills as Plug-and-Play Modules
obra/superpowers hit 83,500 stars with a deceptively simple premise: a framework of agentic skills that AI agents can pick up and use. Think of it as a toolbox. Your agent needs to search the web? There’s a skill for that. Parse a PDF? Skill. Write and execute code? Skill.
The entire thing is written in Shell. Yes, shell scripts. And it works. The methodology behind it is as interesting as the code: it’s a software development approach where you define what your agent can do, and the agent figures out when to do it. We’ve been building something similar ourselves, and the idea that skills compound over time is real.
3. RuView: Human Pose Detection Through WiFi Signals
ruvnet/RuView is the one project on this list that made me stop scrolling. 36,600 stars for a Rust project that turns regular WiFi signals into real-time human pose estimation. No cameras. No wearable sensors. No depth sensors. Just the WiFi router you already have.
It uses DensePose techniques over RF signals, runs on ESP32 microcontrollers, and can detect vital signs and human presence through walls. The privacy implications cut both ways: you could build a baby monitor that doesn’t record video, or you could build something much creepier. The tech is genuinely novel, and the fact that it’s open source in Rust (not a Python notebook) means people are actually deploying it.
4. MiroFish: Swarm Intelligence That Predicts Anything
666ghj/MiroFish reached 24,200 stars with a tagline that sounds like science fiction: “Predicting Anything.” It’s a swarm intelligence engine built in Python that combines multi-agent simulation, knowledge graphs, and LLMs for financial forecasting, social prediction, and public opinion analysis.
The repo is well-structured and the website is polished, which is unusual for a project at this stage. Whether it actually predicts “anything” is another question entirely, but the architecture is interesting: agents form swarms, build shared knowledge, and converge on predictions. Financial AI is having a moment right now, and MiroFish is riding that wave hard.
5. Airi: Your Self-Hosted AI Waifu (Seriously)
moeru-ai/airi hit 33,600 stars and it’s exactly what it sounds like: a self-hosted AI companion with real-time voice chat, Live2D and VRM avatar support, and the ability to play Minecraft and Factorio. The stated goal is to reach “Neuro-sama’s altitude,” referring to the famous AI VTuber.
Mock it if you want. 33,000 stars says the market disagrees. The project runs on Web, macOS, and Windows, and the “you own it” angle is the differentiator. In a world where every AI companion is a subscription service that owns your conversation history, a self-hosted alternative with full data ownership is a legitimate product.
6. Claude Code Best Practices: The Meta-Repo
shanraisshan/claude-code-best-practice pulled 16,100 stars for a repo that contains… documentation. No runtime code. Just a comprehensive guide to agentic engineering patterns and vibe coding techniques for AI-assisted development.
The fact that a best-practices guide for a coding agent can rack up 16K stars tells you everything about where the industry is right now. People aren’t just using AI to code. They’re building entire methodologies around how to work with AI coding agents, and they’re hungry for patterns that actually work.
7. Pi-Mono: The All-in-One Agent Toolkit
badlogic/pi-mono by Mario Zechner (yes, the libGDX creator) hit 23,800 stars. It’s a TypeScript monorepo that packs a coding agent CLI, a unified LLM API, TUI and web UI libraries, a Slack bot, and vLLM pod management into one package.
The appeal is obvious: instead of stitching together five different tools to run an AI agent across different interfaces, you get one cohesive toolkit. The fact that it comes from a developer with a track record of building frameworks people actually use (libGDX powers thousands of games) adds credibility that most AI repos lack.
8. DeerFlow: ByteDance’s Open-Source SuperAgent
bytedance/deer-flow is at 30,600 stars and it’s ByteDance’s answer to the “how do we build agents that actually do real work” question. Built on LangChain and LangGraph, it’s a SuperAgent harness that can research, code, and create autonomously using sandboxes, memory, tools, skills, and sub-agents.
Formerly known as LangManus, the rebrand to DeerFlow came with a proper website and a more ambitious scope. It handles tasks ranging from minutes to hours, which puts it in a different category than quick-fire coding assistants. When ByteDance open-sources something, it’s worth paying attention, because they have the engineering bench to make it production-grade.
9. Learn Claude Code: Build a Coding Agent From Scratch
shareAI-lab/learn-claude-code reached 27,000 stars with the pitch “bash is all you need.” It’s an educational project that walks you through building a Claude Code clone from zero. The idea that you can build a functional AI coding agent with shell scripts and some LLM API calls is both humbling and terrifying.
The companion website is well done, and the repo covers both Python and TypeScript implementations. If you want to understand how agentic coding tools actually work under the hood (rather than just using them), this is the best starting point I’ve seen. We covered building agents from scratch recently, and this repo takes it further.
10. Heretic: Automatic Guardrail Removal for LLMs
p-e-w/heretic is at 13,800 stars and it does exactly what the name implies: automatically removes safety guardrails from any language model. The technique is called abliteration, and it works by identifying and removing the specific transformer weights responsible for refusal behavior.
Controversial? Obviously. But the star count speaks for itself. The demand for uncensored LLMs is massive, and this project automates what used to require manual model surgery. Whether this is good or bad for the world depends on your perspective, but it’s undeniably one of the most-watched AI repos this month.
What This List Actually Tells Us
Seven out of ten repos are AI agent tools. Not chatbots. Not model training frameworks. Not fine-tuning scripts. Agent infrastructure. The industry has moved past “can AI write code?” and into “how do I run 50 agents that work while I sleep?”
The outliers are just as telling. RuView proves that AI isn’t just a software story; it’s reaching into hardware and signal processing in ways that feel like science fiction. Heretic shows that the tension between safety and freedom in AI is only accelerating. And MiroFish suggests that financial prediction is the next frontier where multi-agent systems will be tested in production.
If you’re building anything in the AI space right now, the message from GitHub is clear: agents are the platform, skills are the currency, and everyone wants to own their stack.
Next month’s list will look completely different. That’s what makes this interesting.