GitHub Trending Weekly Digest — March 23-28, 2026

This week's GitHub Trending spotlight reveals a fascinating convergence: AI agents are growing up. From enterprise-grade orchestration frameworks to specialized research assistants, the ecosystem is shifting from simple chatbots to systems that can genuinely work for hours on complex tasks. Let's dive into the 19 standout projects that captured developers' attention.

🔥 Trending Dominators (4+ Days)

DeerFlow — ByteDance's SuperAgent Framework

🔗 github.com/bytedance/deer-flow
Trended: 4 days (Mar 23, 24, 25, 26)

What it does: An open-source SuperAgent framework that handles multi-hour tasks through sandboxes, persistent memory, sub-agents, and a skill system. Think of it as a complete runtime environment for AI agents—file systems, memory management, Docker isolation, and inter-agent coordination all in one package.

Why it matters: Most AI agents can barely handle a 5-minute task before losing context or making mistakes. DeerFlow proves that with the right infrastructure (LangGraph orchestration, persistent storage, real sandboxes), agents can tackle genuine workflows: deep research reports, slide deck generation, multi-step coding projects, even video production pipelines.

Tech: Python, LangGraph/LangChain, Docker/Kubernetes sandboxes, MCP server integration, Telegram/Slack/Feishu gateways, GPT-SoVITS voice cloning, multi-LLM support (OpenAI/Moonshot/Azure/DeepSeek/Gemini/Ollama).

🚀 Repeat Performers (2-3 Days)

last30days-skill — Real-Time Community Intelligence

🔗 github.com/mvanhorn/last30days-skill
Trended: 3 days (Mar 25, 26, 27)

What it does: An AI agent skill that searches Reddit, X (Twitter), YouTube, Hacker News, Polymarket, and web sources for the past 30 days of discussion on any topic, then generates evidence-backed narratives with real citations.

Why it matters: The AI world reinvents itself monthly. This skill keeps you current by aggregating what communities are actually upvoting, sharing, betting on (via Polymarket prediction markets), and debating. It uses multi-signal quality ranking: relevance scoring, bidirectional text similarity, cross-platform convergence detection, and temporal recency decay. V2.9.5 adds Bluesky support and comparison mode ("X vs Y").

Tech: Python, Node.js 22+, Claude Code/Codex CLI/Gemini Extensions, ScrapeCreators API (Reddit/TikTok/Instagram), Twitter GraphQL (Bird client), YouTube transcripts, Polymarket API, Hacker News API, SQLite (watchlist mode), Parallel AI/Brave/OpenRouter web search.

MoneyPrinterV2 — Automated Online Income Workflows

🔗 github.com/FujiwaraChoki/MoneyPrinterV2
Trended: 2 days (Mar 23, 24)

What it does: A Python app that automates online monetization workflows: Twitter bots with scheduled posting, YouTube Shorts automation, Amazon affiliate marketing, local business lead generation, and cold email outreach.

Why it matters: Content creation and marketing take time. MoneyPrinterV2 automates the grind—scheduled Twitter threads, auto-generated YouTube Shorts (KittenTTS voice synthesis + gpt4free), affiliate link injection, business email discovery, and cold outreach sequences. Requires Python 3.12; email outreach needs Go runtime.

Tech: Python 3.12, CRON schedulers, Twitter/YouTube/Amazon Affiliate APIs, Go (email), KittenTTS (TTS), gpt4free (LLM access).

Pascal Editor — Web-Based 3D Architecture Design

🔗 github.com/pascalorg/editor
Trended: 2 days (Mar 24, 25)

What it does: A browser-based 3D building editor built with React Three Fiber and WebGPU. Real-time wall, floor, ceiling, and roof editing with collaborative design support.

Why it matters: Traditional architecture tools (AutoCAD, SketchUp) are expensive and desktop-only. Pascal Editor runs in the browser with WebGPU rendering, MVC architecture for scene data, undo/redo, spatial queries, and CSG boolean operations. Perfect for architects, designers, and developers doing rapid prototyping.

Tech: React 19, Next.js 16, Three.js WebGPU, React Three Fiber, Zustand (state), Zod (validation), three-bvh-csg (boolean geometry), Turborepo, Bun.

Dexter — Autonomous Financial Research Agent

🔗 github.com/virattt/dexter
Trended: 2 days (Mar 26, 28)

What it does: An autonomous agent specialized in deep financial analysis. Thinks, plans, and learns like Claude Code, but for financial research. Turns complex questions into step-by-step research plans, auto-selects tools to gather financial data (income statements, balance sheets, cash flows), and self-reviews its work iteratively.

Why it matters: Financial analysis requires collecting, integrating, and validating data from multiple sources—a time-consuming, error-prone process for humans. Dexter automates it with task planning, autonomous execution, self-verification, and real-time market data access. Built-in loop detection and step limits prevent runaway processes.

Tech: TypeScript, Bun runtime, OpenAI/Anthropic/Google/xAI/OpenRouter, Financial Datasets API (institutional-grade data), Exa/Tavily web search, LangSmith evaluation, WhatsApp gateway integration.

Deep-Live-Cam — Real-Time Face Swap & Video Deepfake

🔗 github.com/hacksider/Deep-Live-Cam
Trended: 2 days (Mar 27, 28)

What it does: Real-time face swapping and one-click video deepfake using a single image. Live camera mode: select face → select webcam → press Live. Supports multi-face simultaneous replacement, mouth mask for accurate articulation, and live movie/performance viewing with any face.

Why it matters: A productivity tool for the AI-generated media industry—helps artists animate custom characters, create engaging content, and prototype clothing designs on models. Built-in checks prevent processing inappropriate media (nudity, graphic violence, war footage).

Tech: Python 3.11, ffmpeg, GFPGANv1.4 (face enhancement), inswapper_128_fp16.onnx (swap core), CUDA Toolkit 12.8.0 (Nvidia GPU), cuDNN v8.9.7, CoreML (Apple Silicon M1/M2/M3), DirectML (Windows), OpenVINO (Intel).

AI-Scientist-v2 — Fully Autonomous Scientific Research

🔗 github.com/SakanaAI/AI-Scientist-v2
Trended: 2 days (Mar 27, 28)

What it does: A workshop-level fully autonomous science system that generates hypotheses, runs experiments, analyzes data, and writes scientific manuscripts. Already produced the first fully AI-written paper accepted by peer review.

Why it matters: Unlike v1 (which relied on human-written templates), v2 generalizes across ML domains with progressive agentic tree search. Uses Best-First Tree Search (BFTS) to explore research paths, supports parallel workflows, debugs failed nodes, and grows multiple independent trees. LLM generates ideas (checked for novelty via Semantic Scholar), agentic tree search executes experiments, and the system auto-analyzes results and drafts papers.

Tech: Python 3.11, PyTorch, CUDA, OpenAI/Gemini/Amazon Bedrock (Claude), Semantic Scholar API (literature + novelty check), LaTeX (chktex), poppler, AIDE tree search framework. Cost: ~$15-20 per experiment (Claude 3.5 Sonnet), ~$5 per paper draft.

Twenty — Modern Open-Source CRM

🔗 github.com/twentyhq/twenty
Trended: 2 days (Mar 27, 28)

What it does: A community-driven, modern Salesforce alternative. Self-hostable CRM with full data ownership, inspired by Notion/Airtable/Linear UX patterns.

Why it matters: Traditional CRMs (Salesforce) are expensive and lock in customer data, forcing renewal fees. Twenty provides open-source, self-hosted CRM with complete data control. Customizable objects/fields, flexible views (filters, sorting, grouping, kanban, tables), custom roles/permissions, workflow automation (triggers/actions), email/calendar/file integrations.

Tech: NestJS + PostgreSQL + Redis + BullMQ (backend), React + Jotai (state) + Linaria (CSS-in-JS) + Lingui (i18n) (frontend), TypeScript + Nx (monorepo), Docker Compose (self-hosting), Chromatic (UI testing), Greptile (code review), Sentry (error tracking), Crowdin (translation).

💡 Single-Day Standouts

Project N.O.M.A.D — Self-Contained Offline Survival Computer

🔗 github.com/Crosstalk-Solutions/project-nomad
Trended: Mar 23

Offline-first knowledge and tool collection for emergency/wilderness/no-network scenarios. Built-in offline Wikipedia (Kiwix), medical references, survival guides, ebooks, local AI chat (Ollama + Qdrant RAG), Khan Academy courses (Kolibri), offline maps (ProtoMaps), crypto tools (CyberChef), notes (FlatNotes). One-click install, browser-based access.

Tech: TypeScript, Docker, Kiwix (offline content), Ollama + Qdrant (local AI), Kolibri (education), ProtoMaps (offline maps), CyberChef (data tools), FlatNotes (notes).

PentAGI — Autonomous Penetration Testing Agent

🔗 github.com/vxcontrol/pentagi
Trended: Mar 23

Fully autonomous AI agent system for complex penetration testing tasks. 20+ professional security tools (nmap, metasploit, sqlmap, etc.), all operations in isolated Docker sandbox. Intelligent memory system (stores historical test results + successful methods), knowledge graph integration (Graphiti + Neo4j), expert delegation system, detailed vulnerability report generation.

Tech: Python, Go + GraphQL (backend API), React + TypeScript (frontend), PostgreSQL + pgvector (vector storage), Neo4j (knowledge graph), Docker (sandbox), Grafana + Prometheus (monitoring), Langfuse (LLM observability), supports OpenAI/Anthropic/Gemini/AWS Bedrock/Ollama/DeepSeek/GLM/Kimi/Qwen.

Browser Use — AI Agent Browser Automation

🔗 github.com/browser-use/browser-use
Trended: Mar 23

Makes websites accessible to AI agents, automates online tasks easily. Provides browser automation for AI agents—operate webpages like humans (click, type, screenshot). Supports job application automation, grocery shopping, PC component selection, etc. CLI + Python API, built-in ChatBrowserUse/Google Gemini/Anthropic Claude support. Browser Use Cloud for production-grade stealth browsers and scalable infrastructure.

Tech: Python, Playwright/Chromium (browser automation), LangChain compatible, custom tool extensions, Claude Code Skill integration.

Supermemory — AI Memory Engine & Context Layer

🔗 github.com/supermemoryai/supermemory
Trended: Mar 24

AI-era memory engine, ranks #1 on LongMemEval, LoCoMo, and ConvoMem benchmarks. Provides persistent memory and user profiling for AI. Auto-learns conversations, extracts facts, builds user profiles, handles knowledge updates/contradictions, auto-forgets outdated info. Supports hybrid search (RAG + Memory), multimodal file processing (PDF, image OCR, video transcription, code AST chunking), Google Drive/Gmail/Notion/OneDrive/GitHub real-time sync.

Tech: TypeScript/Python SDK, hybrid retrieval (vector + keyword + reranking), temporal memory management, user profiling engine (static facts + dynamic context), WebHook real-time sync, multimodal extractors, IndexedDB local persistence.

MoneyPrinter Turbo — AI-Powered Automated Short Video Generator

🔗 github.com/harry0703/MoneyPrinterTurbo
Trended: Mar 24

One-click high-quality short video generation using AI LLMs. Just provide a topic/keyword → auto-generates script, sources materials, adds subtitles, background music, and synthesizes video. Supports vertical 9:16 (1080x1920) and horizontal 16:9 (1920x1080), batch generation. Integrates OpenAI/Moonshot/DeepSeek/Tongyi Qianwen/Google Gemini/Ollama, Azure/Edge TTS (9 realistic voices), Edge/Whisper subtitles. Pexels copyright-free material library (or use local assets). Web UI + API, Docker one-click deployment.

Tech: Python, MVC architecture, ImageMagick (image processing), FFmpeg (video synthesis), Whisper (speech recognition), Edge/Azure TTS, OpenAI/Moonshot/DeepSeek/Gemini API, Streamlit web interface, Docker, GPT-SoVITS voice cloning support.

LiteLLM — Python SDK & AI Gateway

🔗 github.com/BerriAI/litellm
Trended: Mar 25

Python SDK and proxy server (AI Gateway) that calls 100+ LLM APIs in OpenAI format. Supports cost tracking, guardrails, load balancing, and logging.

Why it matters: Enterprises and developers need to use multiple LLM providers (OpenAI, Anthropic, Google, Azure), but each has different API formats, auth methods, and features. LiteLLM provides a unified interface—one codebase to access all LLMs, plus enterprise-grade cost management, access control, and monitoring.

Tech: Python, OpenAI-compatible API, 100+ LLM Providers (Bedrock, Azure, VertexAI, Anthropic, Cohere, etc.), Docker, MCP (Model Context Protocol), A2A Agent Protocol, cost tracking & analysis.

RuFlo — Multi-Agent Orchestration for Claude

🔗 github.com/ruvnet/ruflo
Trended: Mar 25

Leading agent orchestration platform for Claude. Deploy intelligent multi-agent swarms, coordinate autonomous workflows, build conversational AI systems with enterprise architecture, distributed swarm intelligence, RAG integration, and native Claude Code/Codex integration.

Why it matters: Single AI agents struggle with complex enterprise tasks—they need collaboration. RuFlo provides an agent orchestration platform supporting multi-agent coordination, distributed intelligence, RAG (retrieval-augmented generation), native Claude Code support, enabling enterprises to build complex AI workflows and autonomous systems.

Tech: TypeScript, Claude API, Multi-Agent Swarms, RAG (Retrieval-Augmented Generation), Claude Code/Codex Integration, Enterprise Architecture, Distributed Swarm Intelligence.

oh-my-claudecode — Zero-Learning-Curve Multi-Agent Orchestration

🔗 github.com/Yeachan-Heo/oh-my-claudecode
Trended: Mar 26

Multi-agent orchestration framework providing team-first, zero-learning-curve orchestration for Claude Code—no need to learn Claude Code, just use OMC directly.

Why it matters: Simplifies Claude Code's complex multi-agent task orchestration. Provides auto-parallelization, persistent execution, intelligent model routing, cost optimization, and skill learning system. Developers can describe needs in natural language and let the system auto-assign tasks to specialized agents.

Tech: TypeScript, Claude Code plugin system, tmux CLI workers, Codex/Gemini CLI integration, WASM edge modules, OpenClaw Gateway integration, Docker.

RuView — WiFi DensePose (Computer Vision Without Cameras)

🔗 github.com/ruvnet/RuView
Trended: Mar 26

Transforms WiFi signals into real-time human pose estimation, vital sign monitoring, and presence detection—no cameras, no wearables, no internet, just physics.

Why it matters: Provides completely privacy-preserving human sensing technology—analyzes WiFi signal CSI (Channel State Information) perturbations to reconstruct human position, respiration rate, heart rate, and presence. Suitable for medical monitoring, retail analytics, smart buildings, disaster rescue (through-wall search and rescue). All without cameras or wearables.

Tech: Rust (810x performance boost, 54K fps), Python, ESP32-S3 hardware mesh (~$1/node), Docker, WASM, RuVector self-learning vector memory system, attention mechanisms, graph algorithms, Fresnel geometry.

VibeVoice — Open-Source Frontier Speech AI

🔗 github.com/microsoft/VibeVoice
Trended: Mar 27

Microsoft's open-source speech AI model family, including TTS and ASR models. Core innovation: continuous speech tokenizer operating at ultra-low frame rate 7.5 Hz (acoustic + semantic), efficiently preserving audio fidelity while significantly improving computational efficiency for long sequences.

What it does: VibeVoice-ASR-7B can process 60-minute audio in a single pass, generating structured transcripts with "who (speaker), when (timestamp), what (content)," supports custom hotwords. VibeVoice-Realtime-0.5B is a lightweight real-time TTS model, supports streaming text input and robust speech generation up to 10 minutes long, first-heard latency ~300ms.

Tech: Transformers (Hugging Face v5.3.0+ native support), next-token diffusion framework, LLM, diffusion head, vLLM inference acceleration, Qwen2.5 1.5B base model, multilingual support (50+ languages), speaker diarization, timestamp annotation, custom hotwords.

Superpowers — Skills & Methodology for AI Coding Assistants

🔗 github.com/obra/superpowers
Trended: Mar 28

A skill framework and software development methodology for AI programming assistants (Claude Code, Codex, Cursor, OpenCode).

Why it matters: Traditional AI coding assistants often jump straight to writing code, lacking systematic thinking. Superpowers provides a complete workflow: from requirements analysis, design specs, implementation plans, to test-driven development (TDD) and code review. Through a "skill trigger" mechanism, AI assistants automatically follow best practices, avoiding skipped steps and bad code.

Tech: Plugin system (supports Claude Code, Cursor Agent, Codex, OpenCode, Gemini), Git Worktrees (parallel dev branches), Subagent-driven development, Test-Driven Development (enforced RED-GREEN-REFACTOR cycle).

Core skills: brainstorming (pre-design thinking), test-driven-development (write tests first), systematic-debugging (four-stage root cause analysis), subagent-driven-development (parallel multi-agent rapid iteration).

Design philosophy: Test-driven, systematic over ad-hoc, reduce complexity, verify over declare.

📊 This Week's Themes

Agent Infrastructure Maturity

DeerFlow, RuFlo, oh-my-claudecode: Enterprise-grade orchestration with memory, sandboxes, and skill systems
Superpowers: Methodology shift from "write code fast" to "build systematically"

Domain-Specific Agents

Dexter (financial research), PentAGI (penetration testing), AI-Scientist-v2 (scientific discovery)
Specialization beats generalization when you need depth

Memory as First-Class Citizen

Supermemory: #1 on AI memory benchmarks
DeerFlow: Persistent memory across multi-hour tasks
last30days-skill: Temporal memory of community discourse

Offline-First & Privacy-Preserving AI

Project N.O.M.A.D: Complete offline knowledge ecosystem
RuView: WiFi-based sensing without cameras
VibeVoice: Open-source speech AI you can self-host

Content Automation at Scale

MoneyPrinterV2/Turbo: Automated YouTube Shorts + Twitter
Deep-Live-Cam: Real-time face swap for video production
Pascal Editor: Browser-based 3D architecture design

Developer Experience Evolution

LiteLLM: Unified API for 100+ LLMs
Browser Use: Make any website agent-accessible
Twenty: Modern CRM UX inspired by Notion/Airtable

💡 Takeaway

The agent revolution isn't about better chatbots—it's about infrastructure. This week's projects share a common thread: they're building the plumbing that lets AI agents actually work. Memory systems that persist across sessions. Sandboxes that isolate execution. Orchestration layers that coordinate multi-agent teams. Knowledge graphs that enable reasoning over time.

We're witnessing the transition from "AI that answers questions" to "AI that does work." The next frontier isn't smarter models—it's better agent runtimes.

Compiled by Tommy Zhang | March 29, 2026