GitHub Open Source Weekly 2026-03-25: Skills Ecosystem Explodes, Flash-MoE Runs 397B Params on a MacBook, Agent Harness War Heats Up
Data period: 2026-03-17 – 2026-03-25 (Rolling 7 days) Sources: GitHub Trending weekly + monthly, GitHub Search API, HN Algolia
TL;DR: Three storylines unfolded simultaneously this week. First, the Skills ecosystem is exploding after Anthropic opened the standard — 5 out of the top 10 new repos are skills-related, from MiniMax's official tech stack to The Minimalist Entrepreneur author Sahil Lavingia's personal skills. "Packaging knowledge as a skill and sharing it with the community" is now a full-blown trend. Second, danveloper/flash-moe, built by a CVS Health VP of AI using pure C/Metal, runs Qwen3.5-397B-A17B on a MacBook Pro with 48GB RAM and scored 393 points with 121 comments on HN — the week's undisputed community buzz champion. Third, the agent harness race is white-hot: affaan-m/everything-claude-code (+21,490 ★/week) briefly overtook veteran obra/superpowers (+19,621 ★/week), with both now surpassing 100K total stars.
📈 Fastest Growing — Weekly Star Gains Top 10
Source:
github.com/trending?since=weekly🔁 = Also appears in monthly trends (sustained momentum signal)
| # | Project | +Stars/Week | Total Stars | Language | Created |
|---|---|---|---|---|---|
| 1 🔁 | affaan-m/everything-claude-code | +21,490 | 104,819 | JavaScript | 2026-01 |
| 2 🔁 | obra/superpowers | +19,621 | 110,358 | Shell | 2025-10 |
| 3 🔁 | 666ghj/MiroFish | +11,768 | 41,818 | Python | 2025-11 |
| 4 | Crosstalk-Solutions/project-nomad | +10,479 | 15,248 | TypeScript | 2025-06 |
| 5 🔁 | bytedance/deer-flow | +10,201 | 43,085 | Python | 2025-05 |
| 6 | jarrodwatts/claude-hud | +7,069 | 12,626 | JavaScript | 2026-01 |
| 7 | FujiwaraChoki/MoneyPrinterV2 | +6,512 | 24,759 | Python | 2024-02 |
| 8 | TauricResearch/TradingAgents | +6,234 | 40,792 | Python | 2024-12 |
| 9 | unslothai/unsloth | +3,719 | 58,019 | Python | 2023-11 |
| 10 | harry0703/MoneyPrinterTurbo | +1,637 | 52,574 | Python | 2024-03 |
🆕 Top New Repos — This Week's Top 15 Newcomers
Source: GitHub Search API (
created:2026-03-17..2026-03-25, sorted by total stars)
| # | Project | Total Stars | Language | Created |
|---|---|---|---|---|
| 1 | MiniMax-AI/skills | 3,867 | C# | 2026-03-17 |
| 2 | HKUDS/ClawTeam | 3,383 | Python | 2026-03-17 |
| 3 | VoltAgent/awesome-codex-subagents | 2,421 | — | 2026-03-17 |
| 4 | danveloper/flash-moe | 1,847 | Objective-C | 2026-03-18 |
| 5 | dontbesilent2025/dbskill | 1,413 | — | 2026-03-20 |
| 6 | louislva/claude-peers-mcp | 1,109 | TypeScript | 2026-03-21 |
| 7 | math-inc/OpenGauss | 1,076 | Python | 2026-03-19 |
| 8 | lxf746/any-auto-register | 1,065 | Python | 2026-03-18 |
| 9 | zarazhangrui/codebase-to-course | 1,055 | — | 2026-03-22 |
| 10 | slavingia/skills | 1,038 | — | 2026-03-23 |
| 11 | eze-is/web-access | 988 | JavaScript | 2026-03-18 |
| 12 | truongduy2611/app-store-preflight-skills | 922 | — | 2026-03-19 |
| 13 | mattprusak/autoresearch-genealogy | 914 | — | 2026-03-18 |
| 14 | BryanLunduke/DoesItAgeVerify | 882 | — | 2026-03-18 |
| 15 | wangziqi06/724-office | 864 | Python | 2026-03-17 |
This Week's Highlights — Fastest Growing Top 10
📈 #1 — affaan-m/everything-claude-code 🔁 | 10-Month-Polished Agent Harness from an Anthropic Hackathon Winner
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
This week +21,490 ★ | Total ★104,819 | JavaScript | MIT
everything-claude-code started at the Cerebral Valley x Anthropic hackathon (February 2026). Author Affaan M. packaged 10 months of daily heavy Claude Code usage into a complete suite — configs, skills, and hooks. With 1,282 tests, 98% coverage, and 102 static analysis rules, it's one of the most thoroughly documented agent harnesses in the community.
The standout feature is the built-in AgentShield security scanner. In an ecosystem where everyone is sharing skills, scanning your Claude Code config for vulnerabilities, misconfigurations, and injection risks is a pragmatic defense layer. Installation is as simple as adding it via the Claude Code plugin marketplace.
This week's +21,490 stars (104K total) briefly overtook superpowers, signaling that developers want "ready-to-use agent harness packages" more urgently than "learn-from-scratch methodologies."
📈 #2 — obra/superpowers 🔁 | The Agent Harness Methodology Benchmark with 40K Monthly Growth
An agentic skills framework & software development methodology that works.
This week +19,621 ★ | Total ★110,358 | Shell | MIT
superpowers, maintained by Jesse Vincent and the Prime Radiant team, holds the highest total star count among agent harnesses (110K+) and has appeared in monthly trends for multiple consecutive weeks. Its philosophy: agents shouldn't start writing code immediately — they should first formalize the problem into a spec, then produce an implementation plan that "any motivated junior engineer could follow."
The key difference from everything-claude-code is directional: superpowers is methodology-first (spec → plan → TDD red/green), while everything-claude-code is tooling-first (install and go). On HN, contributors have forked it to address amnesia (memory loss), bloat (config sprawl), and safety rails issues, showing enough community activity to produce substantive fork discussions.
📈 #3 — 666ghj/MiroFish 🔁 | Swarm Intelligence Prediction Engine: 10 Days of Vibe-Coding, $4.1M Angel Investment
A Simple and Universal Swarm Intelligence Engine, Predicting Anything.
This week +11,768 ★ | Total ★41,818 | Python | AGPL-3.0
MiroFish has a remarkable backstory: author Guo Hangjiang built the first version in 10 days of vibe-coding. Within 24 hours of posting a rough demo, Shanda Group (founded by former Chinese billionaire Chen Tianqiao) committed $4.1M in angel funding. The core concept is "swarm intelligence prediction" — it scrapes real-world seed information (news, policies, financial signals), automatically constructs a high-fidelity digital world, and lets thousands of personality-driven agents with long-term memory and behavioral logic freely interact, socialize, and evolve, ultimately outputting prediction reports.
Developers have already connected MiroFish to Polymarket trading bots, simulating 2,847 digital personas before each trade, reporting $4,266 in profits across 338 trades. Note the AGPL-3.0 license, which has restrictions on commercial integration.
📈 #4 — Crosstalk-Solutions/project-nomad | Offline AI Knowledge Server for Doomsday Prep
Project N.O.M.A.D, is a self-contained, offline survival computer packed with critical tools, knowledge, and AI to keep you informed and empowered—anytime, anywhere.
This week +10,479 ★ | Total ★15,248 | TypeScript | Apache-2.0
project-nomad takes a completely different approach: it's not about helping you write better code — it's about giving you access to knowledge, AI assistance, and tools when you're off-grid, off-power, or in an emergency. It integrates Ollama local AI Chat (with RAG), Kiwix offline Wikipedia and medical references, Kolibri education platform (Khan Academy courses), ProtoMaps offline maps, and CyberChef data tools.
Against the backdrop of growing global infrastructure reliability concerns, "sovereign computing" tools like this are seeing noticeably increased attention. Requires Ubuntu 22.04+ or Debian 12+; a GPU-equipped device is recommended for full AI functionality.
📈 #5 — bytedance/deer-flow 🔁 | ByteDance's Open Source SuperAgent: Research, Code, and Create in One Place
An open-source SuperAgent harness that researches, codes, and creates. With the help of sandboxes, memories, tools, skill, subagents and message gateway.
This week +10,201 ★ | Total ★43,085 | Python | MIT
DeerFlow 2.0 is ByteDance's late-February 2026 rewrite (sharing no code with v1), positioned as a "batteries-included SuperAgent harness" — not a framework you assemble yourself, but a complete out-of-the-box agent solution. Built on LangGraph + LangChain, it ships with Docker sandbox execution, filesystem access, memory, skills, sub-agent dispatching, and a message gateway, handling complex tasks that take minutes to hours.
It has appeared in monthly trends for multiple consecutive weeks, confirming sustained community maintenance. Compared to everything-claude-code and superpowers, DeerFlow's distinguishing feature is the enterprise-grade infrastructure design (sandboxed execution, production-ready) that comes with corporate backing.
📈 #6 — jarrodwatts/claude-hud | Real-Time Status HUD Panel for Claude Code
A Claude Code plugin that shows what's happening - context usage, active tools, running agents, and todo progress
This week +7,069 ★ | Total ★12,626 | JavaScript | MIT
claude-hud solves a real pain point: when Claude Code is running tasks, you can't easily see how much context window it's using, which tools are executing, how sub-agents are progressing. This plugin provides a HUD panel showing real-time context usage, active tools, running agents, and todo progress.
For users running long agent tasks (rather than short conversations), visibility is a genuine need. On HN, someone built a Codex HUD version based on claude-hud's concept, showing this design pattern is spreading across tools.
📈 #7 — FujiwaraChoki/MoneyPrinterV2 | Veteran Automation Tool for Online Money-Making
Automate the process of making money online.
This week +6,512 ★ | Total ★24,759 | Python | AGPL-3.0
MoneyPrinterV2, created in 2024, automates YouTube/TikTok content generation, Twitter outreach, and more. It re-entered the charts this week, riding the resurgent wave of AI content creation buzz. AGPL-3.0 licensed — evaluate carefully before commercial use.
📈 #8 — TauricResearch/TradingAgents | Multi-Agent LLM Financial Trading Framework
TradingAgents: Multi-Agents LLM Financial Trading Framework
This week +6,234 ★ | Total ★40,792 | Python | Apache-2.0
TradingAgents is a multi-agent LLM-based financial trading research framework backed by an arXiv paper (2412.20138). The +6,234 weekly stars represent solid performance, but a critical caveat: this is a research framework, not a production trading system. Financial decisions carry significant risk, and any automated trading tool requires rigorous risk controls and regulatory compliance assessment.
📈 #9 — unslothai/unsloth | All-in-One Platform for Local Fine-Tuning + Chat UI
Unsloth Studio is a web UI for training and running open models like Qwen, DeepSeek, gpt-oss and Gemma locally.
This week +3,719 ★ | Total ★58,019 | Python | Apache-2.0
Unsloth Studio's new version integrates a web UI, making local fine-tuning no longer a command-line-only affair. It supports training, chat testing, and reinforcement learning fine-tuning for mainstream open-source models including Qwen, DeepSeek, gpt-oss, and Gemma3. On HN, Show HN: Unsloth Studio - Local Fine-tuning, Chat UI garnered 8 points, with discussion centered on "is fine-tuning on a MacBook actually viable?" — echoing this week's Flash-MoE buzz. The boundaries of what "local AI" can do are expanding fast.
📈 #10 — harry0703/MoneyPrinterTurbo | Veteran AI Short-Video Generator
AI-powered one-click HD short video generation
This week +1,637 ★ | Total ★52,574 | Python | MIT
MoneyPrinterTurbo, created in early 2024, is an AI short-video generation tool still receiving steady long-tail traffic this week. Last push was 2025-12-14; currently in maintenance mode.
This Week's Highlights — Top New Repos Top 10
🆕 #1 — MiniMax-AI/skills | MiniMax's Official Claude Code Skills Package
MiniMax's official skills for Claude Code, Cursor, Codex, and OpenCode developers
Total ★3,867 | C# | MIT | Created: 2026-03-17
MiniMax (the prominent AI startup behind MiniMax-M2, M2.1, M2.5 models) released its official skills repo this week, covering full-stack frontend development, backend API design, iOS/macOS development, shader design, and more. Install via claude plugin install minimax-skills.
This is one of the most symbolically significant data points in the week's "Skills ecosystem explosion": AI model companies are now treating "maintaining their own skills package" as part of their ecosystem strategy, rather than waiting for Anthropic to lead. The same skills standard is now compatible across Claude Code, Cursor, Codex CLI, and OpenCode.
🆕 #2 — HKUDS/ClawTeam | One Command Triggers Full Agent Swarm Automation
ClawTeam: Agent Swarm Intelligence (One Command → Full Automation)
Total ★3,383 | Python | MIT | Created: 2026-03-17
ClawTeam comes from HKU's Data Intelligence Lab (HKUDS). Its core design: "a leader agent calls clawteam spawn to create workers, each automatically assigned a git worktree, tmux window, and identity."
The showcase demo: a leader agent coordinating 8 specialized sub-agents across 8 H100 GPUs, autonomously designing experiments, dynamically allocating resources, and integrating breakthroughs across teams. Fully compatible with Claude Code, Codex, OpenClaw, nanobot, and Cursor. For scenarios requiring large task decomposition across multiple parallel agents, ClawTeam offers a more structured solution than manual agent management.
🆕 #3 — VoltAgent/awesome-codex-subagents | Curated List of 130+ Codex Subagents
A collection of 130+ specialized Codex subagents covering a wide range of development use cases.
Total ★2,421 | MIT | Created: 2026-03-17
An awesome-list style repo collecting 130+ Codex subagents for different development scenarios. Compared to skills, subagents are more granular task execution units. For developers looking to quickly find "someone already built an agent for this specific use case," this list is a solid starting point.
🆕 #4 — danveloper/flash-moe | This Week's HN Heat Champion: Running 397B Params on a MacBook
Running a big model on a small laptop
Total ★1,847 | Objective-C | Created: 2026-03-18
This week's HN top score: 393 points, 121 comments
Author Dan Woods (CVS Health VP of AI Platforms) wrote an inference engine in pure C, Objective-C, and hand-tuned Metal shaders that runs Qwen3.5-397B-A17B on a MacBook Pro M3 Max (48GB RAM) at 4.4+ tokens/sec, with tool calling support and production-quality output.
The technical core exploits MoE (Mixture-of-Experts) model characteristics: although the 397B model is massive, each inference only "activates" 4 out of 128 experts — each expert is ~3.9MB, and with the Apple M3 Max SSD's 17.5 GB/s read speed, loading 4 experts takes under 1 millisecond. Non-expert embedding tables and routing matrices (~5.5GB) stay resident in memory. The entire 209GB model lives on SSD, requiring no GPU memory.
The HN discussion centered on: is this an engineering party trick or a genuinely useful workflow? Most highly-upvoted comments agreed: for batch, non-real-time inference tasks (overnight processing, deep research), 5.5 tokens/sec is perfectly usable; for real-time conversation, it's not fluid enough yet. The no-Python, no-framework design also makes it a valuable reference for Metal inference engineering.
🆕 #5 — dontbesilent2025/dbskill | Business Diagnostics Skills for Claude Code
dontbesilent's business diagnostics skills for Claude Code
Total ★1,413 | Created: 2026-03-20
dbskill exemplifies the week's "personal skills" trend: the author packaged their business diagnostics methodology as Claude Code skills for community sharing. Accumulating 1,413 stars in just 5 days reflects strong developer interest in "opinionated domain skills" — not just coding skills, but business and strategy knowledge packaging also has an audience.
🆕 #6 — louislva/claude-peers-mcp | Let Multiple Claude Code Instances Message Each Other
Allow all your Claude Codes to message each other ad-hoc!
Total ★1,109 | TypeScript | Created: 2026-03-21
claude-peers-mcp solves an intriguing multi-agent coordination problem: if you have multiple Claude Code sessions running simultaneously, how do you let them communicate? This MCP server enables real-time messaging between different Claude Code instances, opening up interesting possibilities — like having a writer agent and reviewer agent negotiate directly across terminal windows.
🆕 #7 — math-inc/OpenGauss | Project-Scoped Lean Workflow Orchestrator
Project-scoped Lean workflow orchestrator from Math Inc
Total ★1,076 | Python | MIT | Created: 2026-03-19
OpenGauss is positioned as a workflow orchestrator for Lean (the formal mathematical verification language), supporting the organization and execution of Lean verification tasks within AI coding agent environments. Worth watching for developers doing formal verification or AI-assisted mathematical research.
🆕 #8 — lxf746/any-auto-register | Automated Account Registration Tool
Total ★1,065 | Python | Created: 2026-03-18
This repo has no official description; the name suggests it automates account registration across various services. Exercise caution and evaluate compliance implications before exploring further.
🆕 #9 — zarazhangrui/codebase-to-course | A Skill That Turns Any Codebase into an Interactive HTML Course
A Claude Code skill that turns any codebase into a beautiful, interactive single-page HTML course for non-technical vibe coders.
Total ★1,055 | Created: 2026-03-22
Another compelling skills use case: automatically generating interactive HTML courses from codebases that non-technical users can understand. A time-saver for developer advocates and technical content creators.
🆕 #10 — slavingia/skills | The Minimalist Entrepreneur Author Sahil Lavingia's Skills
Claude Code skills based on The Minimalist Entrepreneur by Sahil Lavingia
Total ★1,038 | Created: 2026-03-23
Gumroad founder and The Minimalist Entrepreneur author Sahil Lavingia released his own Claude Code skills repo, packaging the book's entrepreneurial methodology into executable skills. This is the week's most emblematic case of "skills as a knowledge distribution vehicle": a book's core wisdom can now be directly executed by AI agents.
The HN discussion Skills are quietly becoming the unit of agent knowledge earned 9 points, with the core thesis: skills are quietly becoming the fundamental unit of AI agent knowledge — much like npm packages for the JavaScript ecosystem, technical knowledge, industry know-how, and personal methodologies are all beginning to circulate in skills format.
Monthly Trend Cross-Reference
Repos appearing in both weekly and monthly trends (🔁 marked):
- everything-claude-code (monthly +21,490): Rising agent harness star, backed by Anthropic hackathon pedigree
- obra/superpowers (sustained monthly growth): Steady agent harness methodology benchmark, highest total stars on the board at 110K
- MiroFish (continued monthly growth): Swarm intelligence prediction engine, sustained attention after $4.1M angel backing
- bytedance/deer-flow (monthly stable): ByteDance's open-source SuperAgent, corporate backing inspires long-term maintenance confidence
The significance of monthly trend presence: all four repos have moved beyond "one-time virality," representing sustained developer community attention rather than topic-driven fleeting hype.
This Week's Trend Insights
Skills Ecosystem: From Tools to Knowledge Distribution Infrastructure
The week's most prominent signal is the concentrated explosion of skills-related repos: MiniMax official skills (3,867 ★), VoltAgent/awesome-codex-subagents (2,421 ★), dbskill (1,413 ★), slavingia/skills (1,038 ★), zarazhangrui/codebase-to-course (1,055 ★), truongduy2611/app-store-preflight-skills (922 ★) — 6 out of the top 15 new repos are directly skills-related. This isn't merely a byproduct of "Claude Code is popular." After Anthropic opened the Skills standard, an ecosystem resembling an npm registry is self-organizing: technical knowledge, personal methodologies, and industry know-how are beginning to circulate and share with skills as the unit of exchange.
The Boundaries of Local Inference Are Expanding Rapidly
Flash-MoE running 397B params on a MacBook and Unsloth Studio supporting local fine-tuning both point in the same direction: what your local hardware can do is expanding faster than most people expect. Apple Silicon's SSD bandwidth (17.5 GB/s) and unified memory architecture form the technical foundation of this "local large model" feasibility wave. For privacy-sensitive scenarios or users concerned about inference costs, 2026 is the year to reassess the assumption that "everything should go through a cloud API."
Agent Harness Competition: Plug-and-Play vs. Methodology
The rivalry between everything-claude-code (+21K/week) and obra/superpowers (+19K/week) reflects two distinct developer needs: the former offers "out-of-the-box" value (1,282 tests, security scanning, marketplace install), while the latter provides a "methodology framework" (spec-first, TDD, composable skills). They're not mutually exclusive — many developers actually use components from both repos. The real significance of this competition: "how to make AI agents get things right" has become a first-class problem in software development, no longer an early adopter's toy.
