GitHub Open Source Weekly 2026-04-01: Claude Code Source Leak Breaks GitHub Records, Skills Frameworks Dominate, AI Agent Ecosystem in Full Bloom
Data period: 2026-03-25 to 2026-04-01 (Rolling 7 days) Sources: GitHub Trending weekly + monthly, GitHub Search API, HN Algolia
TL;DR: The biggest event this week was Anthropic accidentally shipping a complete source map inside the Claude Code npm package — 510,000 lines of TypeScript scattered across GitHub within hours. claw-code (a Rust rewrite) set an all-time GitHub record by hitting 50K stars in just 2 hours. Meanwhile, superpowers and everything-claude-code continued holding the top two Fastest Growing spots, and OpenAI officially released a plugin that lets you call Codex from within Claude Code — the AI toolchain coopetition era is here.
Fastest Growing — Weekly Star Gains Top 10
Source:
github.com/trending?since=weekly🔁 = Also appears on monthly trends (sustained momentum signal)
| # | Project | +Stars/week | Total Stars | Language | Created |
|---|---|---|---|---|---|
| 1 | 🔁 affaan-m/everything-claude-code | +19,869 | 129,198 | JavaScript | 2026-01-18 |
| 2 | 🔁 obra/superpowers | +18,021 | 129,363 | Shell | 2025-10-09 |
| 3 | 🔁 bytedance/deer-flow | +13,560 | 55,736 | Python | 2025-05-07 |
| 4 | 🔁 mvanhorn/last30days-skill | +11,933 | 17,096 | Python | 2026-01-23 |
| 5 | microsoft/VibeVoice | +8,327 | 33,817 | Python | 2025-08-25 |
| 6 | 🔁 NousResearch/hermes-agent | +7,539 | 20,977 | Python | 2025-07-22 |
| 7 | Yeachan-Heo/oh-my-claudecode | +6,933 | 20,142 | TypeScript | 2026-01-09 |
| 8 | hacksider/Deep-Live-Cam | +6,693 | 87,182 | Python | 2023-09-24 |
| 9 | Crosstalk-Solutions/project-nomad | +6,272 | 20,855 | TypeScript | 2025-06-24 |
| 10 | 🔁 TauricResearch/TradingAgents | +4,848 | 45,471 | Python | 2024-12-28 |
Top New Repos — Born This Week Top 10
Source: GitHub Search API (
created:2026-03-25..2026-04-01, sorted by total stars)
| # | Project | Total Stars | Language | Created |
|---|---|---|---|---|
| 1 | instructkr/claw-code | 94,240 | Rust | 2026-03-31 |
| 2 | sanbuphy/claude-code-source-code | 9,956 | TypeScript | 2026-03-31 |
| 3 | ChinaSiro/claude-code-sourcemap | 6,713 | TypeScript | 2026-03-31 |
| 4 | openai/codex-plugin-cc | 6,628 | JavaScript | 2026-03-30 |
| 5 | larksuite/cli | 5,770 | Go | 2026-03-25 |
| 6 | claude-code-best/claude-code | 5,678 | TypeScript | 2026-03-31 |
| 7 | Kuberwastaken/claurst | 5,153 | Rust | 2026-03-31 |
| 8 | TheTom/turboquant_plus | 4,474 | Python | 2026-03-25 |
| 9 | titanwings/colleague-skill | 4,339 | Python | 2026-03-30 |
| 10 | magnum6actual/flipoff | 2,567 | JavaScript | 2026-03-26 |
This Week's Spotlight — Fastest Growing Top 10
#1 — affaan-m/everything-claude-code | The Performance Optimization System for Claude Code's Agent Harness
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
This week +19,869 ★ | Total ★129,198 | JavaScript | MIT | Created: 2026-01-18
everything-claude-code held onto the #1 weekly gain spot with +19,869 stars — slightly down from last week's +21,490, but still dominant. The repo positions itself as a cross-tool "agent harness performance optimization system" spanning Claude Code, Codex, Cursor, and other AI coding tools, covering Skills, Instincts, Memory, Security modules with a research-first methodology.
This week's momentum ties directly to the Claude Code source leak: among the 44 leaked feature flags, several capabilities (agent swarms, coordinator mode) closely match harness designs that everything-claude-code has been building, sparking intense community debate. For developers, this repo remains one of the most comprehensive entry points into the Claude Code plugin ecosystem — from memory management to multi-agent coordination.
#2 — obra/superpowers | The Skills Framework That Actually Makes AI Coding Agents Ship Production Code
An agentic skills framework & software development methodology that works.
This week +18,021 ★ | Total ★129,363 | Shell | MIT | Created: 2025-10-09
Built by Jesse Vincent (known as obra), superpowers is the most influential Skills framework in the Claude Code ecosystem, with 129,363 total stars slightly edging out everything-claude-code. Weekly gains came in at #2 (+18,021). Since joining Anthropic's official Claude Code plugin marketplace on January 15, 2026, its credibility has surged.
The core design philosophy is "enforce workflows, don't suggest them" — the agent automatically detects relevant skills before any task and activates them, covering brainstorming, TDD, systematic debugging, and more. It's compatible with Cursor, Codex CLI, OpenCode, Gemini CLI, and other major tools. The HN community debate centers on: compared to pure prompt optimization, is skills-as-workflow the real path to making AI agents reliable?
#3 — bytedance/deer-flow | ByteDance's Open Source SuperAgent for Hours-Long Complex Tasks
An open-source long-horizon SuperAgent harness that researches, codes, and creates.
This week +13,560 ★ | Total ★55,736 | Python | MIT | Created: 2025-05-07
DeerFlow 2.0 is ByteDance's open-source SuperAgent framework that automatically decomposes complex tasks across multiple sub-agents running in Docker sandboxes — meaning it handles tasks that take minutes to hours, not just chat turns. After the 2.0 release on February 28 hit GitHub Trending #1, it's still holding strong at #3 with +13,560 this week.
DeerFlow 2.0 was a complete rewrite, pivoting from a pure research tool to a general-purpose agent orchestration platform supporting OpenAI, Gemini, DeepSeek, Claude, and ByteDance's own Doubao models. For developers looking to self-host an AI research assistant, DeerFlow offers one of the most complete open-source options available — a full stack including memory management, tool calling, and sub-agent scheduling.
#4 — mvanhorn/last30days-skill | AI Research Skill Spanning Reddit, X, HN, and Polymarket
AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web - then synthesizes a grounded summary
This week +11,933 ★ | Total ★17,096 | Python | MIT | Created: 2026-01-23
last30days-skill ranked #4 with +11,933 in weekly gains. This skill positions itself as a "time-aware AI researcher": input any topic, and it simultaneously crawls Reddit, X (Twitter), YouTube, HN, Polymarket, and web search to synthesize a source-backed summary of the latest developments.
The Polymarket integration is particularly noteworthy — it lets the skill query prediction market data, so AI research results include both community sentiment and market probabilities. This adds a valuable dimension for financial analysis and trend forecasting. For developers building their own agent workflows, last30days-skill is a plug-and-play, high-quality "intelligence gathering" module.
#5 — microsoft/VibeVoice | Microsoft's Open Source Voice AI — 90-Minute, 4-Speaker Synthesis
Open-Source Frontier Voice AI
This week +8,327 ★ | Total ★33,817 | Python | MIT | Created: 2025-08-25
VibeVoice is Microsoft's open-source voice AI family covering both TTS (text-to-speech) and ASR (speech recognition). The core technical innovation is an ultra-low 7.5 Hz frame-rate continuous speech tokenizer that dramatically reduces computational overhead for long sequences while maintaining audio quality. The TTS component can synthesize up to 90 minutes of audio with up to 4 different speakers — breaking through the 1-2 speaker limit of most models.
In March 2026, VibeVoice-ASR was integrated into Hugging Face Transformers, marking its entry into the mainstream open-source toolchain. The practical takeaway: this is currently the closest open-source option to commercial services like ElevenLabs under a permissive license (MIT), supporting 50+ languages. This week's momentum is independent of the Claude Code leak — it reflects continued growth in the AI voice space.
#6 — NousResearch/hermes-agent | The Self-Evolving AI Agent That Grows With You
The agent that grows with you
This week +7,539 ★ | Total ★20,977 | Python | MIT | Created: 2025-07-22
Nous Research's hermes-agent is an AI agent that genuinely learns from use: it autonomously builds procedural skills from conversations, continuously optimizes those skills over time, and deepens its understanding of user preferences through Honcho dialectic user modeling. The companion hermes-agent-self-evolution repo implements evolutionary self-improvement using DSPy + GEPA (Genetic Evolution of Prompt Architectures).
The deployment bar is impressively low — runs on a $5 VPS with simultaneous support for Telegram, Discord, Slack, WhatsApp, Signal, and CLI. For developers wanting a "private AI assistant that grows its capabilities over time," hermes-agent offers a more compelling long-term value proposition than static agents. The HN debate centers on: where are the boundaries of self-improvement? Does an agent modifying its own prompts equate to modifying its own behavior?
#7 — Yeachan-Heo/oh-my-claudecode | Teams-First Multi-Agent Orchestrator for Claude Code
Teams-first Multi-agent orchestration for Claude Code
This week +6,933 ★ | Total ★20,142 | TypeScript | MIT | Created: 2026-01-09
oh-my-claudecode's entry point is "Teams-first" — not optimized for individual developers, but designed to let Claude Code run in multi-agent parallel within team environments. It supports running multiple Claude Code sessions concurrently with coordinated cross-session task dispatch, suitable for scenarios requiring large-scale AI agent collaboration (like simultaneously reviewing multiple PRs or developing multiple modules).
This repo has special significance in the wake of the Claude Code leak: the leaked source revealed an "agent swarms" feature flag, and oh-my-claudecode is essentially the community's preemptive implementation before the official feature arrives — demonstrating the open-source community's remarkable ability to anticipate Anthropic's roadmap.
#8 — hacksider/Deep-Live-Cam | Real-Time Face Swap From a Single Photo — Ethics Debate Continues
real time face swap and one-click video deepfake with only a single image
This week +6,693 ★ | Total ★87,182 | Python | AGPL-3.0 | Created: 2023-09-24
Deep-Live-Cam is a perennial trending project, ranking #8 this week with +6,693 and 87,182 total stars marking it as a core reference implementation in the deepfake space. It uses the pre-trained inswapper_128 model, supports NVIDIA CUDA and Apple Silicon, and needs only a single photo for real-time face swapping, including livestream support.
Its reappearance this week likely correlates with escalating discussions around AI-generated content regulation. Notably, the repo has built-in ethical safeguards: the program blocks processing of indecent or sensitive material, and the README explicitly requires users to obtain consent from the person being swapped. For developers: the AGPL-3.0 license means any commercial application built on this must be open-sourced — a deliberate choice that itself serves as an abuse constraint.
#9 — Crosstalk-Solutions/project-nomad | The Offline Survival Computer With Built-In AI + Wikipedia + Maps
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 +6,272 ★ | Total ★20,855 | TypeScript | Apache-2.0 | Created: 2025-06-24
Project NOMAD (Node for Offline Media, Archives, and Data) was built by Chris Sherwood (Crosstalk Solutions) as an offline knowledge system using Docker to integrate local AI (Ollama), 99.6GB of Wikipedia, Khan Academy courses, OpenStreetMap offline maps, and security tools — designed to work permanently without internet after installation.
This repo hit GitHub Trending #1 on March 21 and is still growing strongly. Its audience extends beyond the prepper community — increasing numbers of developers see it as a blueprint for self-hosted AI knowledge bases, or as a practical tool for connectivity-limited environments (boats, remote areas, disaster zones). Minimum hardware: 4GB RAM dual-core CPU, but 500GB SSD + 16GB RAM is recommended for best results.
#10 — TauricResearch/TradingAgents | Multi-Agent LLM Financial Trading Framework
TradingAgents: Multi-Agents LLM Financial Trading Framework
This week +4,848 ★ | Total ★45,471 | Python | Apache-2.0 | Created: 2024-12-28
TradingAgents simulates financial institution roles (researcher, analyst, risk manager, trader) using multiple LLM agents, accompanied by an arXiv paper (2412.20138) explaining its academic foundation. This week's +4,848 shows sustained monthly trend momentum (🔁) — it's been attracting consistent attention since late last year.
For developers interested in quantitative research with AI agents, TradingAgents provides a complete "simulated trading floor": different-role agents debate each other, propose opposing viewpoints, and ultimately reach consensus. HN discussions noted that among Claude Code's leaked feature flags, agent coordination capabilities align naturally with TradingAgents' multi-agent debate architecture.
This Week's Spotlight — Top New Repos Top 10
#1 — instructkr/claw-code | The Fastest Repo to Ever Hit 50K Stars: Claude Code Leak's Rust Rewrite
The fastest repo in history to surpass 50K stars, reaching the milestone in just 2 hours after publication. Better Harness Tools that make real things done. Now writing in Rust using oh-my-codex.
Total ★94,240 | Rust | Created: 2026-03-31
In the early hours of March 31, Chaofan Shou (@Fried_rice, Solayer Labs intern) publicly disclosed on X that Anthropic had accidentally leaked a complete Claude Code source map via npm — 510,000 lines of TypeScript spread across GitHub within hours. claw-code was the fastest derivative: rather than a simple mirror, it immediately announced a Rust rewrite of core functionality and crossed 50K stars within 2 hours of publication — the fastest any repo has ever reached that milestone in GitHub history.
Additional fallout from the leak: Anthropic confirmed that users who installed or updated Claude Code (@anthropic-ai/claude-code v2.1.88) between 2026-03-31 00:21–03:29 UTC may have pulled in a malicious axios version (1.14.1 or 0.30.4) containing a Remote Access Trojan (RAT). If you updated Claude Code during that window, audit your installation environment immediately.
The HN discussion (20 points, 7 replies) centered on: is claw-code's Rust rewrite a pure architectural reference, or does it touch Anthropic's IP? The general community consensus: as long as the effort involves understanding design patterns from the leaked source map rather than directly copying code, the legal risk is manageable.
#2 — sanbuphy/claude-code-source-code | One of the First Mirrors of the Leaked Claude Code Source
It will be revised soon.
Total ★9,956 | TypeScript | Created: 2026-03-31
This is one of the direct mirror repos from the Claude Code leak. The 9,956 stars reflect intense developer curiosity. The fork count (18,687) far exceeds stars, indicating many developers downloaded the full codebase for local study.
What did the leaked source reveal? Beyond the 44 unreleased feature flags (KAIROS always-on Claude, ULTRAPLAN 30-minute remote planning, Buddy companion mode, coordinator mode, agent swarms), developers gained insight into Claude Code's prompt engineering strategies, memory architecture, and security sandbox implementation. For AI coding tool developers, this is a rare "competitor source code learning opportunity."
#3 — ChinaSiro/claude-code-sourcemap | Mirror Archive of the Source Map Itself
Total ★6,713 | TypeScript | Created: 2026-03-31
This repo mirrors the core artifact of the leak: the 59.8 MB .map file itself. Source maps are debug symbol files for browser developer tools. Anthropic exposed it because someone forgot to add *.map to .npmignore.
Technical context: Claude Code uses Bun for bundling, and Bun's bundler generates source maps by default unless explicitly disabled. This seemingly minor configuration oversight resulted in the accidental open-sourcing of the entire codebase. Anthropic has since patched and released a new version.
#4 — openai/codex-plugin-cc | OpenAI Moves Into Claude Code: The Official Codex Plugin
Use Codex from Claude Code to review code or delegate tasks.
Total ★6,628 | JavaScript | Apache-2.0 | Created: 2026-03-30
This is one of the most strategically significant repos of the week. On March 30, OpenAI released an official plugin letting users call Codex directly from within Claude Code, offering three modes: /codex:review (standard code review), /codex:adversarial-review (adversarial review that actively challenges design decisions), and /codex:rescue (full task handoff for Codex to debug or reimplement).
Installation requires a ChatGPT subscription (free tier works) or an OpenAI API key, plus Node.js 18.18+. The strategic implication: OpenAI isn't trying to replace Claude Code — it's deliberately establishing a presence inside a competitor's ecosystem, reaching developers who otherwise wouldn't use OpenAI tools. The trend of developers simultaneously using multiple AI tools from different companies in their actual workflow is accelerating.
#5 — larksuite/cli | Official Lark/Feishu CLI With 19 Built-In AI Agent Skills
The official Lark/Feishu CLI tool, maintained by the larksuite team — built for humans and AI Agents. Covers core business domains including Messenger, Docs, Base, Sheets, Calendar, Mail, Tasks, Meetings, and more, with 200+ commands and 19 AI Agent Skills.
Total ★5,770 | Go | MIT | Created: 2026-03-25
The official Lark/Feishu CLI tool offers 200+ commands covering messaging, documents, spreadsheets, calendar, mail, tasks, meetings, and other core business scenarios, with 19 built-in AI Agent Skills. Following GitHub CLI, this is another major enterprise collaboration platform natively integrating AI agents into its CLI tooling.
For developers and AI agent builders working within the Feishu ecosystem, larksuite/cli provides a standardized interface for directly operating Feishu workspaces without needing to study individual API endpoints. The 19 built-in Skills let AI agents schedule meetings, create documents, and manage tasks directly — this "enterprise API + Skills framework" combination represents an important trend of AI agents entering enterprise workflows.
#6 — claude-code-best/claude-code | The "Enterprise Reliability" Version With TypeScript Type Fixes
原汁原昧 Claude Code 可运行版; Bun 可编译执行版; Typescript 类型全修复; 企业级可靠性
Total ★5,678 | TypeScript | Created: 2026-03-31
This repo positions itself as a "directly bun run dev bootable version" based on the leaked source, focusing on fixing TypeScript type errors and ensuring lock file consistency, emphasizing enterprise-grade reliability over feature expansion. The 5,678 stars indicate a significant number of developers want to locally study or run Claude Code, not just read static code.
#7 — Kuberwastaken/claurst | Rust Rewrite of Claude Code + Architecture Breakdown
Claude Code in Rust & a Breakdown of How it Works
Total ★5,153 | Rust | Created: 2026-03-31
claurst is independent developer Kuberwastaken's Rust rewrite of the Claude Code architecture, accompanied by a detailed architecture analysis blog post. Compared to claw-code, claurst leans more educational — the author treats the rewrite as an opportunity to learn and explain Claude Code's design philosophy, making it ideal for developers who want to deeply understand AI coding agent architecture.
#8 — TheTom/turboquant_plus | Quantitative Trading Toolkit
Total ★4,474 | Python | Apache-2.0 | Created: 2026-03-25
This week saw both TradingAgents (Fastest Growing #10) and turboquant_plus (New Repos #8), reflecting sustained interest in quantitative trading tools within the open-source community. turboquant_plus doesn't have an official description yet, but its technical tags and initial traction of 4,474 stars make it a quant community project worth tracking.
#9 — titanwings/colleague-skill | Turn a Departing Colleague's Knowledge Into a Skill
将冰冷的离别化为温暖的 Skill，欢迎加入数字生命1.0！Transforming cold farewells into warm skills? Welcome to Digital Life 1.0.
Total ★4,339 | Python | MIT | Created: 2026-03-30
This intriguing repo attempts to encapsulate a "colleague's work knowledge" into AI agent skills, allowing organizations to retain operational knowledge after personnel turnover. Part of its 4,339-star initial traction comes from the evocative description, but it touches a real enterprise pain point: how to systematically transfer knowledge from experienced employees into an AI-operable skills format.
#10 — magnum6actual/flipoff | Free Split-Flap Display Emulator
Free split-flap display emulator for any TV. The classic flip-board look, without the $3,500 hardware.
Total ★2,567 | JavaScript | Created: 2026-03-26
The only purely non-AI repo in this week's New Repos Top 10. flipoff uses browser JavaScript to emulate the classic airport/train station split-flap display, turning any TV into a retro flip-board — no $3,500 hardware required. The 2,567 stars show that the nostalgia aesthetic + practical utility combination still has a wide audience.
Monthly Trends Cross-Reference
Five of this week's Fastest Growing Top 10 also appeared on the monthly trends (🔁):
| Project | Monthly Sustained Momentum Significance |
|---|---|
| everything-claude-code | Multi-week Top 2, the benchmark for the Claude Code plugin ecosystem |
| obra/superpowers | Continued user growth since official recognition, the definitive Skills framework |
| bytedance/deer-flow | SuperAgent ecosystem leader, sustained 2.0 release effect |
| mvanhorn/last30days-skill | Standard "researcher" module in the Skills ecosystem |
| TauricResearch/TradingAgents | Long-term reference implementation for the quant AI community |
These five repos' sustained monthly momentum confirms three main threads driving the open-source community: the Claude Code Skills/Harness ecosystem, SuperAgent frameworks, and quantitative AI tools.
This Week's Trend Insights
1. The Claude Code Leak Reshaped the Open Source Landscape
A single event (Anthropic's npm source map oversight) spawned at least 6 distinct GitHub repos within 72 hours (claw-code, claurst, sanbuphy/claude-code-source-code, ChinaSiro/claude-code-sourcemap, claude-code-best/claude-code, tvytlx/ai-agent-deep-dive), and set the all-time fastest star growth record. This wasn't just a security incident — it was the open-source community's collective response to closed-source AI tool architecture being made public. The demand for transparency and reproducibility was demonstrated in the most dramatic way possible.
2. AI Toolchain Coopetition Enters a New Phase: OpenAI Sets Up Camp Inside Claude Code
The release of openai/codex-plugin-cc marks a shift from "walled gardens" to "proactive interoperability" in the AI toolchain: OpenAI isn't waiting for users to switch from Claude Code to Codex — it's directly offering official integration inside a competitor's platform. This strategy mirrors enterprise software's API-first approach: let developers experience your capabilities in their familiar environment first, then pursue deeper migration. Expect similar integrations from other tools (Gemini CLI, Copilot, etc.) to follow.
3. The Skills Ecosystem Moves From Framework Competition to Vertical Specialization
While Skills framework leaders (superpowers, everything-claude-code) continued to dominate in recent weeks, this week's newcomers (last30days-skill, larksuite/cli with 19 built-in Skills, colleague-skill) signal a shift from "my Skills framework is better than yours" to "what Skills does this specific vertical need?" The trends to watch next: enterprise collaboration (Feishu, Notion, Slack) × AI agent Skills combinations, and community demand for knowledge-transfer Skills.
