Awesome Claude Code Power Ups

A vetted, opinionated list of the Claude Code skills, plugins, MCP servers and hooks that are actually worth installing — with star counts verified against the GitHub API, licence and security notes, install commands, and the hype-vs-reality verdict for each.

Awesome

Every entry here was pulled from the firehose of "10 Claude Code plugins you NEED" posts, then checked the boring way: README read in full, hooks and scripts inspected, stars and last-push date pulled from the GitHub API, licence confirmed, and the marketing claim compared with independent measurements where any exist. Tools that failed the check are listed too — in the Hype check section — because knowing what not to install is half the value.

Scope: things that plug into Claude Code (and usually Codex, Cursor and Gemini CLI via the same SKILL.md format). Not a list of every AI repo on GitHub.

---

Contents

• How to read this list
• Quick start: the 20-minute install
• Engineering workflow
• Front-end & design
• Browser, testing & integrations
• Memory, context & token cost
• Decision-making & research
• Marketing & SEO
• Diagrams, video & media
• Reference repos worth bookmarking
• Try with caveats
• Hype check: popular but not worth installing
• How installs work
• Safety checklist before installing anything
• Contributing

---

How to read this list

Field	Meaning
★	GitHub stars from the API at review time, rounded. A big number means viral, not good — several 100k+ repos are in Hype check.
Rating	1–5 for usefulness in a real Claude Code workflow, discounted for licence, security and terms-of-service risk.
Type	skill = markdown instructions (zero execution surface) · plugin = skills + commands + possibly hooks that run code · mcp = a server process · cli = a tool Claude calls
Trigger	auto (Claude picks it when your prompt matches) · slash (/command) · phrase (a trigger sentence)

Ratings are one reviewer's judgment. Open an issue if you disagree — with evidence.

---

Quick start: the 20-minute install

If you only do one thing, install these in this order. All are MIT/Apache, from accountable authors, and five of the seven are plain markdown.

# skills (markdown only, no hooks)
npx skills@latest add addyosmani/agent-skills -g -a claude-code -y --skill '*'
npx skills@latest add vercel-labs/agent-skills -g -a claude-code -y --skill web-design-guidelines react-best-practices
npx skills@latest add emilkowalski/skills -g -a claude-code -y --skill '*'
npx skills@latest add pbakaus/impeccable -g -a claude-code -y --skill '*'
npx skills@latest add shadcn/improve -g -a claude-code -y

# official browser automation
claude mcp add playwright -s user -- npx -y @playwright/mcp@latest      # Windows: prefix the command with  cmd /c
npx playwright install chromium

# diagrams
claude plugin marketplace add cathrynlavery/diagram-design && claude plugin install diagram-design@diagram-design

Then open a fresh claude session and type / to see what arrived.

---

Engineering workflow

Tool	Type · Trigger	What it does	★	Rating
addyosmani/agent-skills	skill · auto + /spec /plan /build /test /review /ship	24 senior-engineer workflow skills by Addy Osmani with a router that picks the right one per task. Ships a controlled comparison against competing packs. MIT.	98k	5
shadcn/improve	skill · /improve quick|deep|security|execute	Your strongest model audits the repo across nine categories and writes self-contained plans to plans/; a cheaper model executes them in a git worktree. Never edits source. MIT.	9k	4.5
DietrichGebert/ponytail	plugin · auto (SessionStart hook) + /ponytail-review	"Write less code" decision ladder: reuse → stdlib → native → dependency → build. The only token-saver here with an independent JetBrains benchmark: −10% cost, −15% code, not the marketed −50%. MIT.	144k	4
HKUDS/CLI-Anything	plugin · /cli-anything <repo>	HKU Data Science lab tool that generates a CLI + SKILL.md for any software with source or an API, so agents drive Blender, OBS, n8n without an MCP. Review generated CLIs before running. Apache-2.0.	50k	4
rebelytics/one-skill-to-rule-them-all	skill · auto (CLAUDE.md line)	task-observer watches sessions and proposes skill improvements for you to approve — never edits on its own. CC BY 4.0.	3k	4
Onome-AJ/security-sweep-plugin	plugin · /security-sweep	Pattern-based OWASP sweep. Tiny, no hooks, unusually honest README ("read patterns.md before trusting this"). Not a SAST replacement. MIT.	<100	3

Front-end & design

Tool	Type · Trigger	What it does	★	Rating
vercel-labs/agent-skills	skill · auto ("review my UI")	web-design-guidelines audits UI against Vercel's published Web Interface Guidelines; react-best-practices has 40 severity-ranked rules. Official.	31k	5
pbakaus/impeccable	skill + hook · /impeccable craft|critique|audit|polish	Paul Bakaus' anti-slop front-end rules plus a zero-API-cost detector that flags 61 generic-UI patterns after every edit. Run /impeccable init per project. Apache-2.0.	70k	4
emilkowalski/skills	skill · auto	Animation, Apple-style UI and library-selection taste from the animations.dev author. MIT.	40k	4
VoltAgent/awesome-design-md	reference · manual	Drop-in DESIGN.md files distilled from Stripe, Airbnb, Revolut, Kraken etc. Copy one into a project root. Plain text, zero risk. MIT.	117k	4
MengTo/threeui · threeui.com	reference · copy prompt	160+ free three.js hero sections and shaders, each with a copy-paste agent prompt. Pro tier exists; free tier is real.	—	4
Leonxlnx/taste-skill	skill · auto	Design references for less generic UI. Largely a remix of the two above; several forks share the name — install this exact repo. MIT.	92k	3.5
beautifului.dev	shadcn registry · per component	20 MIT React/Tailwind v4 primitives for agent UIs (streaming text, approval cards, tool chips). Studio lead magnet; one component pulls a commercial icon set.	<300	3

Browser, testing & integrations

Tool	Type · Trigger	What it does	★	Rating
microsoft/playwright-mcp · playwright-cli	mcp / cli · auto	Official. Claude opens the site it built, screenshots, clicks through flows. Microsoft now points coding agents at the CLI as the token-cheap path. Apache-2.0.	37k / 13k	5
Figma Dev Mode MCP	mcp · auto	Official Figma server; needs the desktop app in Dev Mode.	official	5
tester-army/e2e	cli · npx e2e init	Describe a test in English; an agent drives it once through Playwright, then recorded steps replay with no model calls until the UI changes. Pre-1.0, hosted upsell. Apache-2.0.	9k	3
heygen-com/hyperframes	skill · auto	"Write HTML, render video": deterministic MP4 from HTML timelines, built for agents. Expect HeyGen hosted upsell. Apache-2.0.	52k	4
cloudflare/agentic-inbox	app · self-host	Cloudflare's own Workers-based AI mail client. Apache-2.0. Check activity before adopting.	8k	4

Memory, context & token cost

Tool	Type · Trigger	What it does	★	Rating
thedotmack/claude-mem	plugin · auto	Persistent cross-session memory; injects a ~100-token index at start and pulls detail on demand. Keep optional cloud sync off (CLAUDE_MEM_ONLINE_OPTIN=false). Apache-2.0.	99k	4
headroomlabs-ai/headroom	cli / proxy	Compresses tool output locally; publishes per-task benchmarks (21–57%) instead of one headline. For Claude Code it runs as a local proxy — opt in knowingly. Apache-2.0.	73k	4
Graphify-Labs/graphify	skill + cli · /graphify .	Tree-sitter knowledge graph of a repo (local, deterministic). Install only from the Graphify-Labs org — dozens of identical forks exist. Apache-2.0/MIT.	120k	3
Manavarya09/cost-guardian	plugin · auto	Local per-session cost dashboard, no telemetry. Small and unaudited. MIT.	<100	3
JuliusBrussee/caveman	skill (+ optional proxy)	Terse-output skill is harmless and free. The proxy mode is a local MITM on agent traffic under BSL-1.1 with telemetry on by default — skill only.	107k	3

Decision-making & research

Tool	Type · Trigger	What it does	★	Rating
aiwithremy/claude-skills-llm-council	skill · phrase "council this"	Karpathy-style LLM Council: five advisor personas, blind peer review, chairman verdict that keeps dissent. 5–10× tokens — for real decisions only. MIT.	2k	4
0xNyk/council-of-high-intelligence	skill · slash	The other credible "council" implementation; the name is shared by a dozen one-star clones.	4k	3.5
K-Dense-AI/scientific-agent-skills	skill · auto	177 science skills; the ~18 data/stats ones (EDA, statsmodels, scikit-learn, PyMC, Polars) are worth cherry-picking. Soft funnel to the company's hosted platform. MIT.	48k	3
ai-boost/awesome-harness-engineering	reading list	CC0 curated sources on the scaffolding around agents (OpenAI, Anthropic, Fowler).	5k	3

Marketing & SEO

Tool	Type · Trigger	What it does	★	Rating
AgriciDaniel/claude-seo	plugin · /seo page|technical|schema|geo|audit	25 sub-skills + 18 agents; most commands need no API keys; free Google key unlocks Core Web Vitals; drift monitoring baselines a site and diffs after deploys. 410 tests, no telemetry. MIT.	18k	4
coreyhaines31/marketingskills	skill · auto + /cro /seo-audit /pricing-strategy	~40 marketing playbooks from a CRO agency founder. Advisory, well-structured; README funnels to his agency and two paid sponsors. Install with -a claude-code. MIT.	53k	3.5

Diagrams, video & media

Tool	Type · Trigger	What it does	★	Rating
cathrynlavery/diagram-design	plugin · auto + slash	42 editorial diagram types as self-contained HTML with SVG/PNG export and Mermaid/draw.io re-render. No hooks, no paid tier. MIT.	50k	4
bradautomates/claude-video	plugin · /watch <url>	Downloads, extracts frames and transcribes a video so Claude can answer questions about it. Captions free; Whisper optional. MIT.	18k	3.5

Reference repos worth bookmarking

Nothing to install; clone patterns from them.

• Shubhamsaboo/awesome-llm-apps — 100+ agent/RAG/MCP example apps by a Google Cloud AI PM. Tutorials, not "production-ready". 139k ★
• awesomedata/awesome-public-datasets — topic-sorted open datasets, maintained since 2014. 79k ★
• danny-avila/LibreChat — the mature self-hosted multi-model chat UI with MCP support. 45k ★
• skills.sh — Vercel's "npm for agent skills"; the npx skills add installer used throughout this list. Like npm, anyone can publish: read a SKILL.md before adding it.

---

Try with caveats

Real projects, each with one thing the hype left out.

Tool	The catch	Rating
mksglu/context-mode	"98% less context" is a best case; Elastic Licence v2; passes through gh/aws/kubectl credentials.	3
vectorize-io/hindsight	Serious memory system, but the LongMemEval score is vendor-run, it needs Postgres + an LLM call per capture, and the "use your Claude login" provider is a terms-of-service grey area.	3.5
trycua/cua	Excellent cross-OS computer-use infra (YC). An agent with your mouse acts in every logged-in session — VM or spare account only. Install script is `irm …	iex`.
affaan-m/ECC	277k ★, 68 agents, 293 skills, and hooks that run Prettier + tsc on every edit plus a cost tracker that "emits telemetry markers". Duplicates everything above; $19/seat Pro. Try --profile minimal if curious.	3
tigerless-labs/autoharness	Turns sessions into skills automatically — task-observer without the approval step. Scratch repo first.	3
NangoHQ/nango	1,000+ OAuth connectors; "open source" is Elastic Licence and self-host feature parity is unclear.	3.5
nexu-io/open-design	Big "Claude Design alternative" with an image-to-code skill; its third-party listing pages are ad/affiliate-laden — get it from GitHub only.	3
katipally/openlive	On-device voice that drives Claude Code and approves permissions aloud. Young; read the code before giving it your session.	2

Hype check: popular but not worth installing

The section the "10 plugins you NEED" posts never include. Nothing here is malware; each one fails on terms-of-service risk, a funnel dressed as a tool, a dishonest claim, or being redundant with a better option above. Several have more stars than anything in the recommended tables.

Tool	Why	★
diegosouzapw/OmniRoute	Pools 300 providers' free tiers behind one endpoint; every key routes through it; "1.6B free tokens/month" is the sum of everyone's free tiers; proxying subscription access breaks provider terms; a CVE was referenced in independent coverage.	69k
Panniantong/Agent-Reach	Reads X/Reddit/LinkedIn "without API keys" because it scrapes with your exported login cookies — a documented account-ban pattern. Installs by having the agent fetch and execute a remote script.	96k
multica-ai/andrej-karpathy-skills	219k ★ for one CLAUDE.md with four generic principles; no Karpathy endorsement; promotes the author's product. Ponytail + agent-skills already enforce the same.	219k
mukul975/Anthropic-Cybersecurity-Skills	Not by Anthropic — the README says so. 817 skills including offensive ones; brand-borrowing plus context noise.	34k
volcengine/OpenViking	ByteDance context DB; AGPL, licence key for self-managed, needs an embedding model + VLM. claude-mem covers the use case.	40k
browser-use/browser-use	Fine library, but an LLM-in-the-loop browser is slower and dearer than Playwright MCP for the same jobs; cloud browsers are the product.	118k
midudev/autoskills	Bulk-installs 200+ skills you have not read; CC BY-NC licence; inactive.	7k
lharries/whatsapp-mcp · GOWA	Both use the unofficial whatsmeow library that WhatsApp bans accounts for; the MCP one is stale and its README warns about prompt-injection exfiltration of your chats.	6k / 5k
durchnull/observe	Safe but 0 ★, one author; task-observer does the job with a community.	0
ChatGPTNextWeb/NextChat	Self-hostable client is fine; the "no API key" pitch is the hosted credits reseller.	89k
NVIDIA/Personal-AI-Router	Routes requests between Ollama/LM Studio boxes on a LAN; no GPU pooling; nothing for Claude Code.	2k
"Nemotron Ultra + Ollama = Claude Code free forever"	The model is 550B params (~275 GB VRAM); what runs is Ollama Cloud on starter credits, then per-token billing. Real technique, false framing.	—
mcpmarket.com · VoltAgent/official-mcp-servers	Directories with one-click deploys and paid listings, no vetting. Browse if you like; install from the upstream repo with claude mcp add.	—
originkit.dev	Animated components behind a paywall (10 copies/day free), maker unnamed, code fetched from their server. ThreeUI covers it in the open.	—
Apify "2,000 production APIs" catalogs, "top 10 MCP connectors" Google Docs, "comment X for the link" DMs	Affiliate catalogs and lead magnets. The tools they gate are all searchable in a minute.	—

---

How installs work

Four mechanisms, from safest to most invasive:

1. Skills — markdown files Claude reads. npx skills@latest add owner/repo -g -a claude-code -y. Zero execution surface; inspect with cat ~/.claude/skills/<name>/SKILL.md.
2. Plugins — skills plus commands and sometimes hooks (scripts that run on session start, before/after tool calls). claude plugin marketplace add owner/repo then claude plugin install name@marketplace. Hooks are the thing to read before enabling.
3. MCP servers — separate processes Claude talks to. claude mcp add name -s user -- <command>. On Windows wrap npx commands with cmd /c.
4. Proxies — sit between Claude Code and the API (Headroom, Caveman proxy mode, OmniRoute). Highest trust requirement; only two on this list are worth it, and both are opt-in.

Inside a Claude Code session the same actions are slash commands (/plugin marketplace add …); in a terminal use the claude plugin … form. Open a new session after installing so hooks load.

Windows notes (hard-won): put ~\.local\bin, %APPDATA%\npm and C:\Program Files\Git\bin on PATH; MCP servers launched via npx need cmd /c; if claude plugin marketplace add fails with EBUSY on ~\.claude\plugins\marketplaces, antivirus is scanning the fresh clone — git clone elsewhere and claude plugin marketplace add <local path>.

Safety checklist before installing anything

• Read the SKILL.md. It is instructions your agent will follow with your permissions.
• Does it ship hooks? Read them. Does it ship an MCP server? What credentials can it see?
• Does it proxy your traffic or hold your API keys? Treat as a separate trust decision.
• Does it need you to export login cookies or reuse your subscription login? That is a terms-of-service risk on your account.
• Licence: MIT/Apache are safe; BSL, Elastic, AGPL and CC BY-NC have strings.
• Telemetry default: is there an off switch, and is it documented?
• Stars vs substance: check last push date and open-issue count, not the star count. Several 100k+ repos are in Hype check.
• Install from the upstream repo, never from a directory's "deploy" button.

Contributing

PRs welcome — one tool per PR, using the issue template fields: repo, licence, what it does, how it triggers, any hooks/proxy/credential needs, and evidence for the claim (benchmarks, independent reviews). Ratings are argued, not voted. See CONTRIBUTING.md.

License

CC0 1.0 — do whatever you want with this list. The tools linked keep their own licences.