🇰🇷 한국어 자매 페이지는 발행 후 별도 링크 — Korean sister page added after publish.
TL;DR — In a single week of May 2026 GitHub accumulated ~290k stars and gained ~73k new ones, and not one of the trending Top 10 was a "bigger model." Every slot went to perimeter components for AI agents (memory, code index, on-device TTS, stealth browser, twelve-factor manifesto) and local-first infrastructure. The "ask one model, get one answer" era is quietly giving way to "a team of specialist agents collaborating on the work."
[Image #1 — Hero] GitHub Trending Top 10 of May 2026 — a map of the agent-era perimeter</strong>
file:
2026-05-31-gh10-hero.png
📌 Intro — why "Top 10" is interesting
Who this is for — Indie developers picking new tools for side projects, decision-makers planning internal agent infrastructure, and anyone who wants a single-paragraph summary of "what to look at this month."
GitHub Trending refreshes weekly, but aggregating a month reveals the developer community's priorities at that moment. The signal from May 2026 is unambiguous — none of the Top 10 was a large model itself. All ten are around the model: components, infrastructure, tooling.
The Top 10 cluster into four meta-trends.
Meta-trend | Member repos | Core message |
1. Agent infrastructure | codegraph · agentmemory · 12-factor-agents | Agent context, memory, and architecture principles outweigh the model itself |
2. Vertical specialist Skills | academic-research-skills · ViMax · obra/superpowers | Bundled domain-specific Skills beat a generalist model |
3. On-device / local-first | openhuman · supertonic · RuView | Reclaim data and compute from the cloud back to the device |
4. Automation infrastructure | bun | The foundation that high-throughput agent servers stand on |
Quick glossary:
- Skill: a packaged workflow (prompt + tools + examples) for one task (code review, paper writing) that an agent like Claude Code or Cursor can invoke.
- MCP (Model Context Protocol): a standard for connecting agents to external tools (codebases, DBs, files).
- On-device: running models directly on the user's laptop/phone with no cloud API call.
🥇 Tier 1 — Agent infrastructure (Memory · Index · Architecture)
1. codegraph — semantic index for your codebase
colbymchenry/codegraph · TypeScript · 21,035★ (+14,100/week)
Instead of having a coding agent open files one-by-one to build context for every request, it queries a pre-indexed semantic graph. The pitch: cut token consumption at enterprise scale by an order of magnitude.
When to use: large monorepos where Claude Code / Cursor token costs are biting, or PR review bots whose average response time has crept up.
[Image #2] codegraph — file-by-file exploration vs pre-indexed graph</strong>
file:
2026-05-31-gh10-codegraph.png
2. agentmemory — persistent memory across agents
rohitg00/agentmemory · TypeScript · 17,132★ (+6,900/week)
Hits the biggest weakness of agents head-on: they forget when the session ends. A shared persistent memory pool that multiple clients (Claude, Cursor, Codex) can read and write to — the OSS pattern of "accumulating outcomes," familiar to anyone using memorizer-like tools.
When to use: a single user runs several agents in parallel (IDE coding + CLI debugger + chat) and is tired of re-explaining the same project context and preferences every time.
[Image #3] agentmemory — session forgetting vs shared persistent memory</strong>
file:
2026-05-31-gh10-agentmemory.png
3. 12-factor-agents — twelve-factor principles for production agents
humanlayer/12-factor-agents · TypeScript · 22,078★ (+1,900/week)
A manifesto re-applying Heroku's Twelve-Factor App methodology to LLM-driven agents. "Treat prompts as code, version-control them," "log every tool call," "make partial execution a first-class concern" — twelve principles condensed into a single readable page. The first authoritative engineering discipline for a young field, already a senior's go-to reference.
When to use: right before drafting your team's agent design doc, or when you need an authoritative source to cite in a PR review explaining why a pattern is wrong.
[Image #4] 12-factor-agents — the production-agent manifesto</strong>
file:
2026-05-31-gh10-12factor.png
🥈 Tier 2 — Vertical specialist Skills
4. academic-research-skills — full research workflow as Skills
Imbad0202/academic-research-skills · Python · 20,268★ (+11,600/week)
Packages the entire literature search → writing → peer review → revision loop into Claude Code Skills. The canonical example of "instead of asking a generalist model 'write me a paper,' chain specialist Skills together." Five of the Top 20 trending repos in May contain "skills" in the name — Skill bundles have become a first-class asset class.
When to use: when you or your team regularly produce reports, research, or documentation and want to formalize the model-invocation pattern.
[Image #5] academic-research-skills — four-stage Skill pipeline</strong>
file:
2026-05-31-gh10-research-skills.png
5. ViMax — agentic video generation
HKUDS/ViMax · Python · 7,188★ (+2,700/week)
A multi-agent video pipeline from the University of Hong Kong. Four roles — Director → Screenwriter → Producer → Generator — collaborate sequentially/in parallel to produce a single video. The differentiator: explicitly splitting the creative stages instead of leaning on a single text-to-video model.
When to use: short marketing clips or lecture videos that a one-shot model can't produce and that need stage-by-stage control. Still in the early-adopter phase — experimental rather than rock-solid.
[Image #6] ViMax — Director / Screenwriter / Producer / Generator collaboration</strong>
file:
2026-05-31-gh10-vimax.png
6. superpowers (obra) — composable Skill catalog
obra/superpowers · multi-language · 189,000+★ (top by absolute stars)
A collection of meta-workflows for developer daily life — brainstorming, dispatching parallel agents, systematic debugging, TDD, writing implementation plans — all packaged as Skills. Top-tier by absolute stars in May trending, effectively the reference for "break Skills into small, composable units."
When to use: when first introducing workflow patterns into your agent environment, or as a catalog of "what Skills do others actually use."
[Image #7] superpowers — composable Skill catalog</strong>
file:
2026-05-31-gh10-superpowers.png
🥉 Tier 3 — On-device / local-first
7. openhuman — local-only personal AI assistant
tinyhumansai/openhuman · Rust · 26,795★ (+17,100/week — top weekly growth)
A personal assistant that runs entirely on the user's device, no external APIs. Marketed as "Personal Superintelligence," with the key selling point that data never leaves the device. A direct play on the cloud-first AI backlash — privacy, lock-in, and dependency worries.
When to use: users who handle personal schedules, notes, code, or anything they'd rather not push to the cloud; or environments where the internet is unreliable (travel, abroad, enterprise intranets).
[Image #8] openhuman — fully-local assistant with cloud APIs blocked</strong>
file:
2026-05-31-gh10-openhuman.png
8. supertonic — on-device TTS on Apple Neural Engine
supertone-inc/supertonic · Swift · 9,998★ (+3,600/week)
Korean firm Supertone released a multilingual TTS that runs on Apple Neural Engine via ONNX with zero cloud round-trip. Instantly solves the recurring API-cost pain for iOS/macOS apps that want voice responses but watch OpenAI / ElevenLabs bills pile up.
When to use: an app needs in-product voice and users speak a lot, making per-request API cost a problem. Or when offline operation is required.
[Image #9] supertonic — on-device TTS pipeline on Neural Engine</strong>
file:
2026-05-31-gh10-supertonic.png
9. RuView — WiFi-based, camera-free spatial sensing
ruvnet/RuView · Rust · 65,107★ (+6,800/week — top by total stars)
Tracks a person's position, breathing, and heart rate from subtle WiFi-signal disturbances — no cameras, no microphones. Ships with Home Assistant integration, targeting eldercare squarely. When the whole AI trending list is models and agents, a signal-processing tool that sits next to hardware taking the top spot by total stars is a meaningful signal.
When to use: monitoring an aging parent's home, presence-only triggers in bedrooms/bathrooms where cameras are intrusive, or IoT projects that just need "is a human here" detection.
[Image #10] RuView — WiFi-signal breathing and heart-rate tracking</strong>
file:
2026-05-31-gh10-ruview.png
🏗️ Tier 4 — Automation infrastructure
10. bun — all-in-one JS runtime
oven-sh/bun · Zig/Rust · 92,438★ (+2,000/week)
A longtime trending regular, but in May its value as agent-server infrastructure got fresh attention. Order-of-magnitude faster cold-start than Node, perfect for short-lived agent workers. Built-in bundler/test runner/package manager collapses dependency chains. The "oldest" tool on the list, reheated by new workloads.
When to use: short-lived agent jobs that spin up and die, serverless functions, or any JS project where npm/pnpm install time is a daily pain.
[Image #11] bun — cold-start seesaw vs Node.js</strong>
file:
2026-05-31-gh10-bun.png
🔎 Three cross-cutting patterns
- Large models disappeared from trending: For the whole month of May, no foundational-model repos hit the trending Top 10. The center of gravity moved from model → perimeter.
- Skill bundles became a first-class asset: Five of the top twenty trending repos have "skills" in the name. Claude Code Skills / MCP graduated from a spec to tradeable OSS artifacts.
- "On-device" became the marketing line: openhuman, supertonic, RuView all lead with "no cloud." The 2024 question of "which model" is morphing in 2026 into "where does it run."
🛠️ So what should you actually do
- Personal side project: drop one or two superpowers Skills into your workflow. Five-minute cost, immediate pattern learning.
- Internal agent infra designer: turn the twelve principles from 12-factor-agents into a checklist for the next design doc. Cheap authoritative citation in a young field.
- Monorepo operator: stand up codegraph as a PoC and measure Claude Code's baseline token use.
- App developer: replace your TTS vendor with supertonic to flatten monthly API spend and add offline support.
- IoT / smart home: build one family-safety trigger using RuView — one extra WiFi router and you're done.
📚 Closing — this table will look different in a month
GitHub Trending shifts fast. This is a snapshot dated 2026-05-31, and each repo's star count, language, and trending rationale should be re-measured next month. The value here isn't "which exact ten ranked #1" — it's "why these ten ended up trending together."
References
Surveyed on 2026-05-31 · all star counts and growth figures are same-week trending-page snapshots