Claude Code Directory 2026: the Most Installed Subagents, Commands, Skills, Hooks and MCP Servers

Updated 2026-09-27: first edition, built from the aitmpl.com registry snapshot of 2026-09-26, with every listed file and link checked on 2026-09-27.

Short answer: For developers and software teams already paying for Claude Code, these are the free add-ons other users install most, so the payoff comes out of a subscription you already have. The frontend-design skill leads the entire registry at 47,522 installs, the frontend-developer subagent is second at 39,566, and Context7 leads MCP servers at 14,497. Installing any of them costs nothing; running them spends Claude usage, plus any vendor account an MCP server connects to.

The most installed component of each type

Type	Most installed	Installs
Skill	frontend-design	47,522
Subagent	frontend-developer	39,566
MCP server	context7	14,497
Slash command	generate-tests	11,954
Hook	simple-notifications	10,020

Installs are concentrated. The 85 components ranked on this page are 85 of the 1,765 listed across these five types, and they hold 769,589 of the 1,482,656 installs the registry records for those types, or 51.9 percent.

Jump to: subagents, slash commands, skills, hooks, MCP servers, questions.

What the five component types do

Claude Code is Anthropic’s coding agent, and it runs on Anthropic’s Claude models. Each component type extends it in a different way. The definitions below follow Anthropic’s Claude Code documentation, read 2026-09-27; the folders come from the Claude Code Templates installer source.

• Subagent (--agent, lands in .claude/agents/): a separate Claude worker with its own context window, system prompt and tool permissions. Claude hands it a matching task and gets back a summary.
• Slash command (--command, lands in .claude/commands/): a saved prompt you run by typing /name. Anthropic has merged custom commands into skills, and command files keep working.
• Skill (--skill, lands in .claude/skills/): a folder with a SKILL.md file of instructions, sometimes with scripts. Claude loads it when a task matches, or you run it with /name.
• Hook (--hook, lands in .claude/settings.local.json unless you pick project or user settings): a shell command Claude Code runs at a set moment, such as after a file edit or when Claude stops.
• MCP server (--mcp, lands in .mcp.json): a connection to an outside tool, database or API through the Model Context Protocol.

How to install any of them

Every row below links to the component’s page on aitmpl.com, which shows its exact install command and links to the source file on GitHub. The pattern, from the repository README read 2026-09-27, is one npx command run from your project folder:

npx claude-code-templates@latest --agent development-team/frontend-developer --yes

Swap --agent for --command, --skill, --hook or --mcp and use the category and name shown on the component’s page. The installer writes into the folder it runs in, and --yes skips its confirmation prompts. For MCP servers, Anthropic’s own route is the claude mcp add command. Read a hook or MCP config before you install it: both can run code on your machine.

How this list is ranked

Ranked by public install counts, which measure popularity, not quality. Each table sorts one component type by the exact install figure in the registry snapshot, in descending order, with no editorial reordering and no paid placement.

The registry’s installer logs one anonymous event per component it installs and skips logging when CCT_NO_TRACKING or CCT_NO_ANALYTICS is set or CI is true, so CI runs and files copied by hand from GitHub are not in these numbers. We did not benchmark the components. Each note comes from reading the component’s own file, plus the npm and PyPI records for the MCP packages, on 2026-09-27; where a file does something you should know before installing it, the note says so.

Install counts from the open Claude Code Templates registry (aitmpl.com), snapshot 2026-09-26.

The 20 most installed Claude Code subagents

Development roles fill the list: 11 of the 20 come from the registry’s development-team and development-tools categories. Three names, prompt-engineer, context-manager and devops-engineer, appear twice in the registry under different categories; each row here is the more installed copy, and its link goes to that exact file. A subagent’s requests count toward the same Claude usage limits as your main conversation, per Anthropic’s docs, so add the few you will actually delegate to.

Rank	Name	What it does and when to use it	Installs	Link
1	frontend-developer	Builds complete React, Vue or Angular front ends; reach for it when a feature is mostly UI work.	39,566	View
2	code-reviewer	Reviews pull requests and diffs for security holes, code quality and maintainability; run it before you merge.	30,065	View
3	ui-ux-designer	Critiques screenshots, mockups, CSS and design tokens for usability and accessibility; it has no edit tools, so it advises only.	26,117	View
4	backend-architect	Plans services, API style (REST, gRPC or GraphQL), schemas and scaling, and leaves implementation to other agents.	22,995	View
5	fullstack-developer	Builds one feature through database, API and front end in a single pass, such as a complete sign-up flow.	17,534	View
6	prompt-engineer	Designs, tests and tightens prompts for Claude or other models in production, where accuracy and token spend both matter.	15,559	View
7	debugger	Works from crash logs and stack traces to a root cause and a fix; its file pins the claude-sonnet-4-5 model.	13,013	View
8	database-architect	Designs data models and picks database technology for new systems, and hands Postgres tuning to other agents.	11,762	View
9	python-pro	Writes typed, async, tested Python for APIs and tooling; a fit for FastAPI and SQLAlchemy work.	11,390	View
10	context-manager	Keeps shared context in order across multi-agent or long-running sessions, so each agent starts informed.	10,340	View
11	security-auditor	Audits systems and code against security and compliance controls such as SOC 2; read-only, so it reports without editing.	10,199	View
12	test-engineer	Plans test strategy, automates test suites and checks coverage, including tests that run in CI.	9,374	View
13	typescript-pro	Handles advanced TypeScript: complex generics, conditional types and end-to-end type safety across a stack.	9,339	View
14	architect-review	Reviews changes for architectural consistency, SOLID principles and layering; it installs under the name architect-reviewer.	8,722	View
15	mobile-developer	Builds cross-platform apps in React Native or Flutter, with native features and offline data sync.	7,476	View
16	ai-engineer	Generalist for AI features: model choice, RAG, tool use, evals and serving, for classical ML or LLM apps.	7,457	View
17	error-detective	Correlates errors across logs and services to explain cascading failures; useful in the middle of an incident.	7,324	View
18	devops-engineer	Sets up CI/CD pipelines, infrastructure provisioning and monitoring for deployments.	7,155	View
19	documentation-expert	Writes and maintains project documentation, including docs generated from the code itself.	6,713	View
20	task-decomposition-expert	Breaks a large goal into a dependency-ordered work plan and hands pieces to specialist agents; it cannot edit files.	6,694	View

The 20 most installed Claude Code slash commands

Ranks 15 to 19 are one Git Flow set, feature, finish, release, hotfix and flow-status, built around a develop branch; they only pay off if your team already works in Git Flow. Anthropic’s docs, read 2026-09-27, file custom commands under skills, and command files like these keep working.

Rank	Name	What it does and when to use it	Installs	Link
1	generate-tests	Writes a complete test file for a source file or component you name.	11,954	View
2	ultra-think	Structured deep analysis that surfaces assumptions, tests competing options against each other and ends with a confidence-rated recommendation.	8,241	View
3	create-architecture-documentation	Produces architecture documentation with diagrams and architecture decision records.	5,999	View
4	code-review	On-demand review of code quality, security, performance and architecture.	5,416	View
5	refactor-code	Guided refactor that checks test coverage first and writes missing tests before it changes any code.	4,807	View
6	commit	Writes conventional commits with emoji; unless you pass -no-verify it first runs pnpm lint, build and generate:docs.	4,052	View
7	update-docs	Brings project docs back in line with implementation status and API changes.	3,781	View
8	architecture-review	Reviews design patterns across a codebase and returns improvement recommendations.	3,348	View
9	todo	Keeps a todos.md checklist at the project root: add, complete, remove and list tasks, with optional due dates.	2,671	View
10	explain-code	Explains what code does, from overall purpose down to line-by-line detail; handy in an unfamiliar codebase.	2,438	View
11	create-prd	Drafts a product requirements document for a new feature.	2,389	View
12	workflow-orchestrator	Defines, schedules and monitors multi-step automation workflows with task dependencies.	2,287	View
13	add-changelog	Creates or updates a changelog in the Keep a Changelog format.	1,978	View
14	generate-api-documentation	Generates API reference documentation in several output formats.	1,799	View
15	feature	Starts a Git Flow feature branch from develop with consistent naming and tracking.	1,787	View
16	finish	Finishes the current Git Flow branch: features merge into develop; releases and hotfixes merge into main and develop with a tag.	1,681	View
17	release	Opens a Git Flow release branch from develop with a version bump and changelog.	1,636	View
18	hotfix	Opens a Git Flow hotfix branch from main for an urgent production fix.	1,578	View
19	flow-status	Shows Git Flow state: branch type, sync status, pending changes and merge targets.	1,494	View
20	security-audit	Runs an on-demand security and vulnerability review of the codebase.	1,411	View

The 20 most installed Claude Code skills

Of the 20, 12 come from the registry’s development category and 5 from creative-design. The code-reviewer name ranks second here with 31,614 installs and second among subagents with 30,065; they are separate components, so pick the form you want. Five skill folders in this table ship their own license file: four are Apache 2.0, and docx is Anthropic’s proprietary skill, so we describe it by function only.

Rank	Name	What it does and when to use it	Installs	Link
1	frontend-design	Steers new or reworked UI toward a deliberate visual direction and typography rather than template defaults.	47,522	View
2	code-reviewer	Code review toolkit for TypeScript, JavaScript, Python, Swift, Kotlin and Go, with checklists and security scanning.	31,614	View
3	senior-frontend	React, Next.js, TypeScript and Tailwind guidance with component scaffolding and bundle analysis.	28,398	View
4	ui-ux-pro-max	Design reference of UI styles, color palettes, font pairings and chart types across web and mobile stacks.	26,502	View
5	senior-backend	Backend patterns for Node, Express, Go, Python, Postgres, GraphQL and REST, from API scaffolds to query tuning.	24,579	View
6	senior-architect	System design support: architecture diagrams, stack decisions, trade-off analysis and integration patterns.	20,090	View
7	skill-creator	Helps you write, edit and benchmark your own skills, including tuning the description that triggers them.	16,680	View
8	ui-design-system	Builds design tokens, component documentation and developer handoff material for a consistent design system.	16,083	View
9	react-best-practices	Vercel Engineering’s React and Next.js performance rules for request waterfalls, bundle size and rendering; MIT licensed.	12,893	View
10	senior-fullstack	Full-stack scaffolding and code-quality checks for React, Next.js, Node, GraphQL and Postgres projects.	11,906	View
11	seo-optimizer	On-page and technical SEO help: keyword research, meta tags, schema markup and Core Web Vitals.	10,545	View
12	webapp-testing	Tests local web apps through Playwright: checks behavior, captures screenshots and reads browser logs.	10,230	View
13	brainstorming	Makes Claude explore intent and requirements before it builds; its description tells Claude to use it before any creative work.	10,128	View
14	canvas-design	Creates static visual pieces such as posters, saved as PNG or PDF files.	10,123	View
15	mobile-design	Mobile-first design principles for iOS and Android: touch interaction, performance patterns and platform conventions.	9,035	View
16	senior-security	Security engineering help: threat modeling, penetration testing, cryptography and audit preparation.	8,982	View
17	senior-prompt-engineer	Prompt patterns, structured outputs, RAG and agent design for teams building LLM products.	8,747	View
18	git-commit-helper	Reads your staged diff and drafts a descriptive commit message.	6,850	View
19	clean-code	Short coding-standards guide: concise, direct code with no over-engineering and no needless comments.	6,368	View
20	docx	Creates, reads and edits Word .docx files. Proprietary Anthropic license; read its LICENSE.txt before use.	6,136	View

The 10 most installed Claude Code hooks

Hooks fire on events inside a Claude Code session, and Anthropic’s hooks reference warns that command hooks run with your full user permissions. Read our file check below the table before installing any of these.

Rank	Name	What it does and when to use it	Installs	Link
1	simple-notifications	Pops a desktop alert when Claude finishes or needs your input, on macOS or Linux; useful when you step away during long runs.	10,020	View
2	smart-commit	Stages and commits each file Claude edits or creates, one commit per change. Reads CLAUDE_TOOL_FILE_PATH.	3,629	View
3	update-search-year	Adds the current year to web searches that lack one. Returns modifiedToolInput, a field the hooks reference does not list.	3,255	View
4	lint-on-save	Runs ESLint, Pylint or RuboCop on each file Claude edits, by file type. Reads CLAUDE_TOOL_FILE_PATH.	3,177	View
5	smart-formatting	Formats each edited file with Prettier, Black, gofmt, rustfmt or php-cs-fixer, by file type. Reads CLAUDE_TOOL_FILE_PATH.	2,899	View
6	dependency-checker	When Claude edits a dependency manifest, runs npm audit, safety or cargo audit where installed. Reads CLAUDE_TOOL_FILE_PATH.	2,255	View
7	nextjs-code-quality-enforcer	Checks each edited JS or TS file against Next.js App Router conventions and hands problems back to Claude; needs jq.	2,168	View
8	security-scanner	Runs semgrep, bandit and gitleaks where installed after each edit, and flags likely hard-coded secrets. Reads CLAUDE_TOOL_FILE_PATH.	2,161	View
9	run-tests-after-changes	Runs npm run test:quick after every edit and reports the result; needs that script in package.json.	1,895	View
10	file-protection	Meant to stop edits to system paths, production config and vendored folders, but it exits with code 1, which does not block. Reads CLAUDE_TOOL_FILE_PATH.	1,607	View

The 15 most installed MCP servers for Claude Code

Three of the 15 point at packages with a problem on the public registries, checked 2026-09-27: npm marks the packages behind postgresql-integration and github-integration as no longer supported, and the npm package web-fetch launches returned not found. The reference fetch server is published on PyPI as mcp-server-fetch by Anthropic. Anthropic’s MCP docs also advise you to verify you trust each server before connecting it.

Rank	Name	What it does and when to use it	Installs	Link
1	context7	Upstash’s server that feeds version-specific library docs and code examples into Claude’s context; the template expects a Context7 API key.	14,497	View
2	memory-integration	Runs the memory server from the Model Context Protocol project’s own repository, so Claude can store facts and recall them later.	7,814	View
3	playwright-mcp-server	ExecuteAutomation’s community Playwright server: browse pages, take screenshots, run JavaScript and generate test code.	6,123	View
4	postgresql-integration	Lets Claude query a PostgreSQL database from a connection string; npm marks the package it installs as no longer supported.	4,818	View
5	web-fetch	Meant to pull web pages and API responses into Claude; the npm package it launches returned not found on 2026-09-27.	4,733	View
6	github-integration	Repos, issues and pull requests through the GitHub API with a personal access token; npm marks its package as no longer supported.	4,134	View
7	supabase	Supabase’s own server for one project; the template starts it read-only and needs a personal access token.	4,095	View
8	chrome-devtools	The Chrome DevTools team’s server for driving Chrome, debugging pages and analyzing performance.	4,053	View
9	filesystem-access	The filesystem server from the Model Context Protocol project’s own repository; it works only inside the directories you pass it.	3,056	View
10	browser-use-mcp-server	Cobrowser’s community server for driving a browser with the Browser Use library; the template sets an OpenAI API key.	2,328	View
11	playwright-mcp	Microsoft’s Playwright server, which reads pages through accessibility snapshots instead of screenshots.	1,925	View
12	mysql-integration	Lets Claude query a MySQL database from a connection string, through the mcp-server-mysql Python package.	1,584	View
13	deepgraph-nextjs	Code-graph server pointed at the vercel/next.js repository, for questions about Next.js internals rather than your own app.	1,517	View
14	figma-dev-mode	Connects Claude to Figma’s Dev Mode server running on your own machine, for design context while you code.	1,483	View
15	markitdown	Converts PDF, Word, Excel, image and audio files to Markdown, from a Docker image named markitdown-mcp.	1,411	View

Further down the same list are servers for tools we review: Notion (#23, 750 installs), Bright Data (#25, 616), Firecrawl (#27, 592), Browserbase (#35, 415), ElevenLabs (#36, 413), n8n (#42, 279) and Apify (#72, 119). All of them are on aitmpl.com’s MCP list.

Questions people ask about Claude Code add-ons

What is the difference between Claude Code subagents, skills and slash commands?

A subagent is a separate Claude worker with its own context window, system prompt and tool permissions: Claude delegates a task to it and gets a summary back, which keeps noisy work out of your main conversation. A skill is a SKILL.md file of instructions that Claude loads when a task matches its description, or that you run by typing /skill-name. A slash command is the older one-file version of the same idea; Anthropic’s docs, read 2026-09-27, say custom commands have been merged into skills. Use a subagent to isolate a job and a skill to reuse a procedure.

Are Claude Code slash commands deprecated?

They are folded in, not removed. Anthropic’s skills documentation, read 2026-09-27, says a file at .claude/commands/deploy.md and a skill at .claude/skills/deploy/SKILL.md both create /deploy and work the same way, and that existing command files keep working. The docs point new work toward skills because a skill folder can carry supporting files. The installer used on this page still writes commands to .claude/commands/.

Where are Claude Code agents and skills stored?

Project copies live in the repository’s .claude folder: subagents in .claude/agents/, skills in .claude/skills/ with one folder and SKILL.md file per skill, and commands in .claude/commands/. Personal subagents and skills that follow you into every project go in ~/.claude/agents/ and ~/.claude/skills/. The Claude Code Templates installer writes to the project folder you run it in: subagents, skills and commands under .claude, MCP servers into .mcp.json, and hooks into .claude/settings.local.json unless you choose project or user settings.

Are Claude Code skills free?

Installing them is. The Claude Code Templates repository is MIT licensed and its installer is a free npx command. Individual skills can carry their own terms: five of the 20 skills in our table ship a license file, four of them Apache 2.0, while docx is Anthropic’s proprietary skill, whose use is governed by your agreement with Anthropic. Running any component spends Claude usage, though Anthropic’s docs note that a skill’s body loads only when it is used, so a long skill costs almost nothing until you need it.

Should I use a skill or an MCP server?

They solve different problems. An MCP server connects Claude Code to an outside system, such as a database, a browser or an API, so Claude can read and act on it directly. A skill tells Claude how to do a job. Anthropic’s docs, read 2026-09-27, draw the line this way: connect a server when you keep copying data into chat from another tool, and create a skill when you keep pasting the same instructions. The docs also say to verify you trust each server before connecting it.

What are Claude Code hooks used for, and are they safe to install?

Hooks run shell commands at set points in a session, such as after Claude edits a file or when it stops to wait for you, which makes them the place for formatting, linting, notifications and guardrails. Anthropic’s hooks reference says command hooks run with your full user permissions, so read every hook before you install it. When we checked the 10 most installed hooks on 2026-09-27, six depended on a variable Anthropic’s references do not list, and the file-protection guard exits with code 1, which does not block an edit.

Sources

1. Claude Code Templates registry, aitmpl.com, snapshot 2026-09-26.
2. davila7/claude-code-templates on GitHub, README, MIT license and installer source at commit fcf666b, read 2026-09-27.
3. Anthropic, Create custom subagents, read 2026-09-27.
4. Anthropic, Extend Claude with skills, read 2026-09-27.
5. Anthropic, Hooks reference, read 2026-09-27.
6. Anthropic, Environment variables, read 2026-09-27.
7. Anthropic, Connect Claude Code to tools via MCP, read 2026-09-27.
8. npm registry record, @modelcontextprotocol/server-github, read 2026-09-27.
9. npm registry record, @modelcontextprotocol/server-postgres, read 2026-09-27.
10. PyPI, mcp-server-fetch, read 2026-09-27.
11. PyPI, browser-use-mcp-server, read 2026-09-27.