GitHub Trending Weekly Digest — May 18 to May 23, 2026

This week's GitHub Trending reveals a pivotal moment in AI-assisted software development. The ecosystem is maturing rapidly—moving beyond experimental demos to production-grade infrastructure, quality-control frameworks, and specialized domain tooling. From AI coding agents that understand your entire codebase to WiFi-based spatial intelligence, here's what dominated developer attention from May 18-23, 2026.

🔥 Trending Champion (4 Days)

Claude Plugins Official

🔗 github.com/anthropics/claude-plugins-official
Trending Days: May 19, 21, 22, 23

What it does:
Anthropic's curated marketplace for high-quality Claude Code plugins. The repository splits into two sections: /plugins (official Anthropic-developed tools) and /external_plugins (vetted third-party contributions). Every plugin undergoes quality and security review before listing.

Why it matters:
The launch of an official plugin directory signals Claude Code's evolution from experimental tool to production platform. By providing a trusted discovery mechanism and standardized installation flow (/plugin install {name}@claude-plugins-official), Anthropic is building the foundation for a thriving extension ecosystem—much like VS Code's marketplace transformed code editing.

Tech:
Standard plugin structure (.claude-plugin/plugin.json metadata, .mcp.json for MCP server config), supports slash commands, agent definitions, and skills. Third-party developers submit through a formal review process, ensuring baseline quality and safety.

💪 Multi-Day Trending (2-3 Days)

OpenHuman

🔗 github.com/tinyhumansai/openhuman
Trending Days: May 18, 19, 20
Stars (May 18): +1,302

What it does:
Your personal AI superintendent—private, simple, dangerously powerful. OpenHuman connects to 118+ services (Gmail, Notion, GitHub, Slack, Stripe, etc.) via OAuth, auto-fetches data every 20 minutes, and builds a local Memory Tree + Obsidian-compatible wiki. The AI agent gets your full context from day one—no "cold start" period.

Why it matters:
Most AI assistants require days of manual feeding or passive learning. OpenHuman flips the script: it actively pulls your inbox, calendar, code repos, and documents into a searchable knowledge graph stored locally in SQLite. Add a talking desktop mascot that joins Google Meet as a real participant, and you have the most ambitious "agent-as-companion" project in the wild.

Tech:
TypeScript desktop app (Tauri), Memory Tree with ≤3k-token Markdown summaries, TokenJuice compression (80% cost reduction), native voice (STT + ElevenLabs TTS), optional local AI via Ollama. MIT licensed, runs entirely on your machine.

Academic Research Skills

🔗 github.com/Imbad0202/academic-research-skills
Trending Days: May 18, 19, 20
Stars (May 18): +1,302
Stars (May 19): +1,439
Stars (May 20): +3,164

What it does:
End-to-end academic research pipeline for Claude Code: from research question → literature review → paper writing → peer review → revision → publication. Includes 13-agent research team, 12-agent writing flow, 7-agent multi-perspective review (EIC + 3 reviewers + devil's advocate).

Why it matters:
AI won't replace researchers, but it can handle the grunt work—finding references, formatting citations, cross-checking data integrity. ARS introduces Stage 2.5/4.5 integrity gates, citation audit channels (v3.8), and cross-model validation. It's not autopilot; it's a research co-pilot with built-in guardrails against fabrication and plagiarism.

Tech:
Python + Pandoc (DOCX) + LaTeX (APA 7.0/IEEE/Chicago) → PDF via tectonic. Semantic Scholar API for citation verification, VLM chart validation, anti-leakage protocols. Supports PRISMA systematic reviews and Socratic guidance mode.

CodeGraph

🔗 github.com/colbymchenry/codegraph
Trending Days: May 20, 22, 23
Stars (May 20): +1,850
Stars (May 22): +4,294 (total: 15,366)
Stars (May 23): +3,684

What it does:
Pre-indexed code knowledge graph for Claude Code, Codex, Cursor, and OpenCode. Builds a searchable graph of symbols, call chains, and code structure using Tree-sitter + SQLite FTS5. AI agents query the graph instead of scanning files.

Why it matters:
Benchmarked on 7 real-world repos, CodeGraph saves 35% cost, 59% tokens, 49% time, and 70% tool calls on average. On VS Code's TypeScript codebase, it reduced 52 tool calls to 3—from 1m37s to 17s (82% faster, 94% fewer calls). It's the difference between an agent reading every file and an agent using an index.

Tech:
TypeScript, SQLite with full-text search, Tree-sitter AST parsing, native OS file watchers (FSEvents/inotify) for auto-sync. Recognizes 14 web frameworks (Django, Flask, Express, Rails, etc.) and links URL patterns to handlers. 100% local, no API keys.

CLI-Anything

🔗 github.com/HKUDS/CLI-Anything
Trending Days: May 18, 19
Stars (May 19): +1,049

What it does:
"Make every software natively usable by AI agents." One command (/cli-anything <app-path>) auto-generates a complete CLI for any software—preserving 100% professional capabilities, zero compromises. Supports 18 apps (GIMP, Blender, LibreOffice, OBS, etc.) with 2,280+ passing tests.

Why it matters:
AI agents excel at reasoning but fail at using professional tools. UI automation is brittle, APIs are limited, and reimplementations lose features. CLI-Anything generates Python Click CLIs that call the real software (Blender's bpy, LibreOffice headless, ffmpeg) and produce valid project files (ODF, MLT XML, SVG). It's the bridge between AI logic and professional creative tools.

Tech:
7-stage pipeline (analyze → design → implement → test → document → publish), Python Click framework, REPL mode + JSON output for agents, Tree-sitter parsing, ONNX Runtime for model inference.

Superpowers

🔗 github.com/obra/superpowers
Trending Days: May 19, 21

What it does:
A complete software development methodology framework for AI coding agents. Takes you from Socratic requirements refinement (brainstorming) → isolated Git worktrees → detailed task plans (2-5 min granularity) → subagent-driven development → TDD with RED-GREEN-REFACTOR → code review → merge.

Why it matters:
Most AI coding assistants jump straight to writing code. Superpowers enforces a disciplined workflow: plan before coding, test before shipping, review before merging. The subagent-driven development skill spawns child agents for each task and runs two-stage review (spec compliance + code quality). It's how you get AI agents to work for hours without derailing.

Tech:
Skill-based workflow automation, supports Claude Code/Codex/Cursor/Copilot/Gemini/OpenCode. Core principles: TDD (tests first), YAGNI (you aren't gonna need it), DRY (don't repeat yourself), systematic debugging over ad-hoc fixes.

Andrej Karpathy Skills

🔗 github.com/multica-ai/andrej-karpathy-skills
Trending Days: May 20, 21

What it does:
A single CLAUDE.md file that fixes LLM coding pitfalls identified by Andrej Karpathy: making unverified assumptions, over-complicating code, and editing unrelated comments/code.

Why it matters:
Karpathy's May 2026 tweet crystallized three LLM failure modes. This project turns his observations into executable instructions: Think Before Coding (state assumptions, ask when uncertain), Simplicity First (no speculative features), Surgical Changes (only touch what's required), Goal-Driven Execution (define success criteria, loop until verified).

Tech:
Plain-text instruction file, installable as Claude Code plugin or per-project CLAUDE.md. Cursor bundles it as .cursor/rules/karpathy-guidelines.mdc.

AI Engineering from Scratch

🔗 github.com/rohitg00/ai-engineering-from-scratch
Trending Days: May 22, 23
Stars (May 22): +1,333 (total: 11,212)

What it does:
A 435-lesson, 20-phase course (~320 hours) teaching AI engineering from linear algebra to autonomous multi-agent systems. Every algorithm is built from raw math first, then reimplemented in PyTorch/sklearn. Each lesson produces a reusable tool (prompt/skill/agent/MCP server).

Why it matters:
84% of students use AI tools, but only 18% feel professionally ready. Most AI education is fragmented—a paper here, a fine-tuning post there, a cool agent demo somewhere else. This course is the spine: one end is linear algebra, the other is autonomous swarms. When PyTorch appears, you already know what it's doing under the hood.

Tech:
Python, TypeScript, Rust, Julia. Six-step rhythm per lesson: MOTTO (core idea) → PROBLEM (pain point) → CONCEPT (intuition) → BUILD IT (raw math) → USE IT (framework) → SHIP IT (tool). Smart skills: /find-your-level (placement quiz), /check-understanding (per-phase tests).

🎯 One-Day Standouts

Scientific Agent Skills

🔗 github.com/K-Dense-AI/scientific-agent-skills
Trending Day: May 18

What it does:
135 production-ready AI agent skills for scientific research—biology, chemistry, medicine, machine learning, geospatial science. Unified database-lookup skill accesses 78 public databases (PubChem, UniProt, COSMIC, ClinicalTrials.gov, etc.). 70+ Python package skills (RDKit, Scanpy, PyTorch Lightning, BioPython, etc.). 9 integration skills (Benchling, DNAnexus, LatchBio).

Why it matters:
Lowers the barrier for cross-disciplinary research. AI agents can now query ChEMBL for EGFR inhibitors → analyze SAR with RDKit → run virtual screening with DiffDock—all through standardized skills. Includes multi-omics integration (RNA-seq, proteomics, metabolomics) and precision medicine workflows (variant annotation, clinical trial matching).

Tech:
Agent Skills standard (agentskills.io), supports Claude Code/Codex/Cursor/Copilot. Hugging Science integration covers 17 scientific domains (astronomy, biology, chemistry, climate, genomics, materials science, medicine, physics).

Supertonic

🔗 github.com/supertone-inc/supertonic
Trending Day: May 18
Stars: +827

What it does:
Lightning-fast, offline, multilingual TTS via ONNX. 99M parameters (vs. 0.7B-2B for competitors), 44.1kHz output, runs on Raspberry Pi and e-readers in airplane mode. Supports 31 languages, 10 inline emotion tags (<laugh>, <sigh>).

Why it matters:
Privacy-first TTS that works offline. Tested on Raspberry Pi and Onyx Boox Go 6 e-reader—average RTF 0.3× in flight mode. Competes with VoxCPM2/OmniVoice/Qwen3-TTS on benchmarks but 7-20× smaller. Browser extension TLDRL converts any webpage to audio in under 1 second.

Tech:
ONNX Runtime, Python/Node.js/Browser (WebGPU)/Java/C++/C#/Go/Swift/iOS/Rust/Flutter. 99M open-weight model, no upsampler needed.

.NET Skills

🔗 github.com/dotnet/skills
Trending Day: May 21
Stars: +96

What it does:
.NET team's curated collection of core skills and custom agents for AI coding assistants. Covers .NET development, Entity Framework, performance diagnostics, MSBuild, framework migration, MAUI, AI/ML integration, ASP.NET Core.

Why it matters:
Brings professional .NET knowledge to AI agents. Includes dashboard tracking accuracy and efficiency trends. Plugins: dotnet (core), dotnet-data (EF), dotnet-diag (profiling), dotnet-msbuild (build), dotnet-upgrade (migration), dotnet-maui, dotnet-ai (LLM/RAG/MCP), dotnet-aspnet.

Tech:
Agent Skills standard, supports Claude Code/Codex/Cursor/Copilot CLI. Install via /plugin install @dotnet-agent-skills.

RuView

🔗 github.com/ruvnet/RuView
Trending Day: May 22

What it does:
Turns ordinary WiFi signals into real-time spatial intelligence, vital signs monitoring, and presence detection—completely camera-free. Uses WiFi CSI (Channel State Information) to detect breathing (6-30 BPM), heart rate (40-120 BPM), 17-keypoint pose estimation, fall detection (<200ms), and through-wall presence (~5m range).

Why it matters:
Privacy-compliant monitoring without cameras (GDPR/HIPAA-friendly). ESP32-S3 costs $9, full Cognitum Seed system ~$140. Use cases: elderly care, hospital monitoring, retail foot traffic, HVAC optimization, search & rescue (through debris/smoke/walls). 105 edge modules (~400 KB signed binaries) for health, safety, building, retail, industrial, and research.

Tech:
ESP32-S3 hardware, WiFi CSI + multi-band fusion + multi-view attention, RuVector (attention + graph algorithms + compression), self-supervised contrastive learning. Adapts to environment in <30 seconds. Raspberry Pi 5 cold start: 8.4ms for pose estimation.

Chrome DevTools MCP

🔗 github.com/ChromeDevTools/chrome-devtools-mcp
Trending Day: May 22
Stars: +151 (total: 40,731)

What it does:
Chrome DevTools for coding agents. MCP server that gives AI assistants full control over Chrome: navigation, input automation (click/drag/fill/hover/type), performance tracing, network analysis, console logs, screenshots, Lighthouse audits, heap snapshots, extensions.

Why it matters:
Reliable browser automation with DevTools-level depth. Puppeteer handles auto-wait; agents get performance traces, network request details, source-mapped console logs. 43 tools across input (10), navigation (6), performance (3), network (2), debugging (8), memory (5), extensions (5), WebMCP (2).

Tech:
Model-Context-Protocol server, Puppeteer engine, supports standalone (auto-launch isolated Chrome) or connect mode (attach to running Chrome). Slim mode: 3 basic tools (navigate/script/screenshot) for lightweight tasks. Experimental: pageId routing (concurrent agent sessions), visual coordinates (computer-use models), screencast (requires ffmpeg).

Understand-Anything

🔗 github.com/Lum1104/Understand-Anything
Trending Day: May 23
Stars: +1,393

What it does:
Converts codebases, knowledge bases, or docs into interactive knowledge graphs. Multi-agent pipeline analyzes projects and builds a graph of every file, function, class, and dependency. Interactive dashboard for visual exploration.

Why it matters:
New engineers facing 200k-line codebases don't know where to start. Traditional code reading is blind and inefficient. Understand-Anything provides a "teaching graph" that shows how every part fits together—not just complexity, but comprehension. Auto-groups into layers (API, Service, Data, UI, Utility), supports 12 coding patterns (generics, closures, decorators).

Tech:
TypeScript/Python/Rust/Julia, multi-agent pipeline (project scanner, file analyzer, architecture analyzer, tour builder, graph reviewer, domain analyzer). Saves knowledge graph as JSON (.understand-anything/knowledge-graph.json). Supports 14+ platforms (Claude Code, Cursor, VS Code + Copilot, Codex, OpenCode, Gemini CLI, Pi Agent). Multilingual output (--language zh for Chinese).

FinceptTerminal

🔗 github.com/Fincept-Corporation/FinceptTerminal
Trending Day: May 23

What it does:
Native C++20 financial analysis desktop app. Multi-asset analysis (DCF, portfolio optimization, VaR, Sharpe, derivatives pricing), 37 AI agents (Buffett/Graham/Lynch/Munger frameworks + economic/geopolitical), 100+ data connectors (DBnomics, Polygon, Kraken, Yahoo Finance, FRED, IMF, World Bank), real-time trading (crypto WebSocket, stocks, algos), 16 broker integrations, QuantLib suite (18 modules), global intelligence (maritime tracking, geopolitics, satellite data), visual workflow (node editor for automation), AI quant lab (ML, factor discovery, HFT, RL trading).

Why it matters:
Professional-grade buy-side analyst toolkit without Bloomberg-level pricing. Single native binary, no Electron overhead. Terminal-level performance. Dual license: AGPL-3.0 (free for personal/academic) + commercial license required for business use.

Tech:
Pure C++20, Qt6 UI, embedded Python for analysis, QuantLib, supports local LLMs (Ollama) + cloud (OpenAI/Anthropic/Gemini/Groq/DeepSeek). Optional Adanos market sentiment (Reddit/X/news/Polymarket). CMake 3.27.7, Ninja 1.11.1, MSVC 19.38 / GCC 12.3 / Apple Clang 15.0, Python 3.11.9.

🔮 Weekly Themes

1. AI Coding Agent Infrastructure Maturation
Four of the top five trending projects focus on making AI coding agents more powerful: official plugin marketplace (Claude Plugins), pre-indexed code knowledge (CodeGraph), software-to-CLI generation (CLI-Anything), and methodology frameworks (Superpowers). The ecosystem is shifting from "can AI write code?" to "how do we build reliable, efficient, disciplined AI coding workflows?"

2. Quality Control as First-Class Concern
Academic Research Skills introduces integrity gates and citation audits. Andrej Karpathy Skills codifies anti-patterns. Superpowers enforces TDD and code review. The community is learning that unconstrained AI generates fast but fragile code—guardrails are essential.

3. Local-First AI
OpenHuman's Memory Tree, Supertonic's offline TTS, CodeGraph's local-only indexing, and FinceptTerminal's Ollama support all emphasize privacy and independence from cloud APIs. The pendulum is swinging back toward data sovereignty.

4. Domain-Specific Tooling
Scientific Agent Skills (135 research skills), .NET Skills (framework-specific), Chrome DevTools MCP (browser automation), FinceptTerminal (finance)—specialized tooling is replacing generic assistants. The future is vertical.

5. Knowledge Graphs Everywhere
CodeGraph, Understand-Anything, OpenHuman's Memory Tree—all build searchable graphs instead of linear file scans. Graphs compress context, enable instant queries, and reduce token burn. The pattern is clear: AI agents need structured knowledge, not raw text dumps.

💡 Takeaway

This week marks a turning point. AI coding agents are no longer experimental toys—they're becoming production infrastructure. We're seeing official marketplaces, performance benchmarks, methodology frameworks, and domain specialization. The open-source community is building the rails for reliable, efficient, ethical AI-assisted development.

The race is on: Who will build the best agent knowledge graph? Who will standardize the plugin ecosystem? Who will crack multi-agent orchestration? The answers are emerging in public, on GitHub, one PR at a time.

Compiled by Tommy Zhang | May 24, 2026