Best Claude GitHub Repositories — June 2026

The most-starred agent harness in this guide isn't the one ranked first. Neither is the most-starred design skill. If you're choosing Claude Code tooling by star count alone, you're optimizing for the wrong number — adoption measures hype, not whether the thing actually works for you.

This edition runs every finalist through the same filter: real adoption, independent security signals where they exist, the credibility of the author, and — where it exists — a documented head-to-head against the alternative. Two tools earned a spot in more than one category because they genuinely solve more than one problem; those are flagged rather than padded into separate filler entries. Two categories from past editions — video editing and curated directories — are sitting out this round until the field matures.

Star counts below reflect GitHub data as of June 16, 2026, and they move daily. Where an independent security scan flagged something, it's stated directly rather than buried in a footnote.

Agent Harnesses & Multi-Agent Orchestration

These aren't skill packs — they're complete operating environments Claude Code runs inside, replacing the default reactive loop with something closer to a structured engineering process.

ECC has 38K more stars than Superpowers, and Superpowers is still ranked first here — because the criteria that predict whether a harness works for you aren't the criteria that predict whether it goes viral.

Superpowers turns Claude Code from a reactive code generator into something that behaves like a senior developer who plans before writing. It enforces a seven-phase workflow — Brainstorm → Spec → Plan → TDD → Subagent Dev → Review → Finalize — and the defining feature is mandatory test-driven development: Claude deletes code written before tests exist and restarts. It was accepted into the official Anthropic Claude Code plugin marketplace in January 2026, and runs across Claude Code, Codex, Cursor, Gemini CLI, and GitHub Copilot CLI.

Best for: developers who want structured, repeatable workflows with strict TDD and subagent delegation — the default choice for shipping non-trivial features without Claude going off the rails.

Install: /plugin install superpowers@claude-plugins-official

Its Brainstorm → Spec → Plan phase is a complete spec-driven workflow on its own — see the Spec-Driven Development section below.

Everything Claude Code (ECC) is the most comprehensive agent harness available, and the most-starred repo in this category. It ships 135 agents, security scanning via the ecc-agentshield npm package, memory optimization, continuous-learning modules, and production-ready configs. Version 1.9 added selective-install architecture, model routing through NanoClaw v2, and support for 12 language ecosystems. One installation works across Claude Code, Codex, OpenCode, and Cursor — it won the Anthropic Hackathon.

Best for: enterprise teams deploying Claude Code across multiple projects who need security gates and agent orchestration with monitoring at scale.

Install: npm install -g ecc-universal && ecc-agentshield init

"The agent that grows with you." Hermes lives on your server rather than inside an IDE session — it remembers what it learns and gets more capable the longer it runs, through a closed loop: persistent memory with FTS5 search and LLM summarization, autonomous skill creation after complex tasks, and skills that self-improve with use. It talks over Telegram, Discord, Slack, WhatsApp, Signal, and CLI from one gateway, and it's MCP-compatible — so it can expose its sessions and memories directly to Claude Desktop, Cursor, or VS Code. It passed an independent security audit with no high-severity issues flagged.

Best for: anyone who wants an always-on personal assistant that learns continuously, not just a coding-session tool. Worth knowing this isn't a Claude Code replacement — it's a parallel, model-agnostic agent that happens to talk to Claude well.

Spec-Driven Development

Repos that make Claude work from a plan instead of vibes — parsing specs, generating ordered task lists, tracking dependencies, executing in sequence. These two aren't rivals: Spec Kit defines the spec, Taskmaster turns it into a tracked, dependency-aware task queue. Most teams that get serious about spec-first work end up running both.

GitHub's own toolkit for spec-driven development, and the most mature option in the category. The workflow is four slash commands: /speckit.specify → /speckit.plan → /speckit.tasks → /speckit.implement. Specifications become living, executable artifacts instead of throwaway docs — when something doesn't make sense mid-build, you return to the spec instead of guessing. Version 0.10 replaced the legacy --ai flags with a full --integration system plus extensions, presets, and per-event hooks, and now supports 30+ coding agents including Claude Code, Copilot, Cursor, Gemini CLI, Windsurf, and Codex CLI.

Best for: teams that don't standardize on one coding agent — Spec Kit gives everyone a shared workflow regardless of which tool they personally prefer.

Install: uvx --from git+https://github.com/github/spec-kit.git specify init my-project --integration claude

Taskmaster is the project-management layer AI-driven development was missing. Feed it a PRD or feature spec and it breaks the work into ordered tasks with dependency tracking, then exposes 36 MCP tools your coding agent can call directly. It detects and integrates with 13 IDEs — Cursor, Claude Code, Windsurf, VS Code, Roo, Kiro, Zed, and more. The loop: write a spec → Taskmaster generates tasks → Claude Code executes them in order. A recent release added watch mode for real-time task monitoring.

Best for: pairing with Spec Kit rather than replacing it — Spec Kit defines the spec, Taskmaster turns it into a tracked, dependency-aware task queue.

Install: npm install -g task-master-ai && task-master init

Worth restating: Superpowers' own Brainstorm → Spec → Plan phase (above) is a full spec-driven methodology before a single subagent touches code — if you're already running it for the TDD enforcement, you don't need a separate spec tool on top.

Behavioral Rules & Guardrails

Configs that fix Claude's most common failure modes: confident wrong assumptions, scope creep, and hallucinated context.

A single CLAUDE.md file encoding four behavioral rules based on Andrej Karpathy's January 2026 viral observations about AI coding agents. Developer Forrest Chang shipped it the day after Karpathy's post. The four rules: Think Before Coding (no silent assumptions), Simplicity First (no over-engineering), Surgical Changes (don't touch code you weren't asked to touch), and Goal-Driven Execution (verify against success criteria, not steps completed). The personal repo and the multica-ai organization mirror carry identical content — pick whichever, the file is the same.

Best for: any developer who's watched Claude silently rewrite an entire codebase when asked to fix one bug. The fastest legitimate quality improvement available, at zero setup cost.

Install: copy CLAUDE.md into your project root.

Graphify's behavioral angle: by compressing a folder of code, docs, or media into a queryable knowledge graph, it stops Claude from confabulating context it never actually read. The skill is honest about what it found versus what it guessed — a direct guardrail against the overconfident-wrong-assumption behavior Karpathy's rules above try to fix from the prompting side. It passed an independent security scan with no high-severity issues.

Best for: pairing with the Karpathy rules above — one fixes Claude's behavior, the other fixes what Claude actually has accurate information about.

Caution on the #3 pick in this category, amanattar/caveman-claude-skill: positioned by Graphify's own author as a direct comparison point, it takes a simplicity-first approach to forcing plainer, less confidently-wrong output. We couldn't independently confirm a star count or security scan at the time of writing — treat it as a promising community pick to test against Graphify yourself, not a verified flagship.

Memory & Context

Claude forgets everything between sessions by default. These give it structured, persistent memory instead.

Cross-listed from Behavioral Rules above, on its own terms here: Graphify turns any folder — code, SQL schemas, PDFs, screenshots, whiteboard photos, video, audio — into a persistent, queryable knowledge graph instead of a one-shot context dump. It's built around the same workflow Karpathy himself uses with a personal /raw folder of papers, tweets, and notes. The headline number is 71.5x fewer tokens per query versus reading raw files, because every later query reads the compact graph instead of re-ingesting source material. It supports 20 languages via tree-sitter AST, plus Whisper-transcribed video and Claude-vision-extracted images.

Best for: anyone whose project has outgrown what fits in a context window — large codebases, research corpora, or mixed-media knowledge bases that need to persist and stay queryable across sessions.

Install: git clone https://github.com/safishamsi/graphify && cd graphify && uv sync --all-extras

Built by kepano — Obsidian's actual creator and lead — which makes this the most credible single source for combining Claude with Obsidian. It teaches your agent to use the official Obsidian CLI and open formats including Markdown, Bases, and JSON Canvas, following the Agent Skills specification shared across Claude Code, Codex, and OpenCode. Drop it into a .claude folder at the root of your vault.

Best for: people who want a lighter, officially-maintained skill set rather than the deepest possible feature surface — start here before reaching for the community alternative below.

Install: /plugin marketplace add kepano/obsidian-skills && /plugin install obsidian@obsidian-skills

A third pick worth knowing: pablo-mano/Obsidian-CLI-skill goes deeper than the official skill above — full vault control via the Obsidian CLI (v1.12+), covering daily notes, full-text search, task/tag/link querying, and plugin and sync management, not just basic note operations. It ships a .claude-plugin/marketplace.json manifest for direct Claude Code marketplace compatibility — worth it if you want Claude to automate your vault deeply rather than just read and write notes.

Copywriting, SEO & Marketing

Structured Claude skills for content production, SEO audits, and email — built to work inside the agent loop, not standalone prompts.

Built by Addy Osmani of the Google Chrome engineering team, which makes this the highest-credibility pick in the category by a wide margin. It brings the same rigor Google's own Lighthouse audits use to performance, Core Web Vitals (LCP, INP, CLS), WCAG accessibility, SEO crawlability and structured data, and modern best practices. Every skill includes a documented table of common excuses agents use to skip verification steps — with counter-arguments built in — and ends with hard evidence requirements: tests passing, build output, runtime data. "Seems right" is explicitly not sufficient.

Best for: teams who want web-quality audits Claude can actually act on, not just describe — especially anyone shipping content where Core Web Vitals affect real search rankings.

The deepest pure technical-SEO suite in this category: 30+ skills and 10 specialist agents covering full site audits up to 500 pages, technical checks (crawlability, Core Web Vitals, IndexNow), schema markup detection and generation, E-E-A-T content quality assessment, GEO optimization for AI Overviews and ChatGPT search, backlink analysis, hreflang validation for multilingual sites, programmatic SEO, and local SEO with Google Maps intelligence.

Best for: sites with multilingual or programmatic content that need audit depth beyond a generalist skill — relevant if you're running properties in more than one language.

A companion pick from the same author, AgriciDaniel/claude-blog, handles the content-production side: brief-to-draft pipelines, brand voice consistency, keyword integration without stuffing, internal link suggestions, and structured Markdown or HTML output. The natural pairing is claude-seo audits, claude-blog produces. For a narrower need, irinabuht12-oss/email-campaigns-claude is a focused, single-purpose skill for email sequences, subject lines, and send-cadence logic — a smaller footprint than the picks above, but the scope is exactly what it advertises.

UI/UX & Design

Skills that turn Claude into a design-system generator — but the most-starred option isn't automatically the best output. We tested before ranking.

Anthropic's own design skill, included with Claude — no install required. In a documented hands-on comparison (the same prompt fed to both this skill and the much-more-starred third-party alternative below, across a SaaS landing page, a sushi restaurant site, and a hot spring inn site), the official skill held its own against a competitor with over ten times its GitHub star count. Stars measure adoption, not necessarily output quality — this is the clearest example in the entire guide.

Best for: the starting point for any design task. Test it before reaching for a third-party alternative, since you already have it.

Install: claude plugin add anthropic/frontend-design

The world's most popular community design skill by star count, and a genuine reasoning engine rather than a static style guide: 67 UI styles, 161 color palettes, 57 font pairings, 99 UX guidelines, 25 chart types, and 161 reasoning rules that analyze your project and generate a complete design system. It supports Claude Code, Cursor, Windsurf, Antigravity, Codex CLI, Continue, Gemini CLI, OpenCode, and more.

Best for: projects needing a very wide style range — glassmorphism through brutalism — or non-web platforms like Flutter, SwiftUI, and React Native that the official skill doesn't target directly. The caveat that matters most: in that documented head-to-head, ~91.8K stars and a vastly larger preset library still didn't reliably beat the free option you already have installed.

The third pick, Leonxlnx/taste-skill (~37.4K ★), takes a different approach entirely: 11 specialized design variants plus 3 image-generation skills, built around a parameter system that works like an audio equalizer on the AI's design output — sliders such as DESIGN_VARIANCE (1–10, clean centered layouts to asymmetric modern compositions) and MOTION_INTENSITY (1–10, simple hovers to magnetic scroll-triggered effects). It's up 181% in adoption versus March 2026, the fastest-growing third-party design skill we tracked this round — worth A/B testing directly against UI/UX Pro Max for your specific brief.

The Bottom Line

The Superpowers-versus-ECC pair at the top of this list is the whole lesson in miniature: ECC has 38K more stars, and Superpowers is still the #1 pick, because official-plugin-marketplace acceptance and mandatory TDD enforcement matter more for most teams than feature count. The frontend-design-versus-ui-ux-pro-max pair makes the same point from the other direction — ten times the stars, and the free built-in option held its own on identical prompts.

Star count is an adoption signal, not a quality signal. Use it to find candidates, not to make the final call — run your own prompt through the top two or three picks in whichever category matches your problem, and let the output decide.

This guide updates monthly. Video editing and curated directories return next edition.

---

If you're evaluating Claude Code tooling because you're already deep into agentic workflows, our breakdown of Claude Fable 5 covers what changed in the underlying model these skills run on top of.