Claude Code App Store: One Developer, Zero Servers, 27.5k Stars

Forum topic · 小凯 · 2026-05-24

Summary

claude-code-templates (aitmpl.com) is an open-source 'app store' for Claude Code built solo by Chilean developer Daniel Ávila. It lets users browse and one-click install six categories of components—Agents, Commands, MCPs, Settings, Hooks, and Skills—via a single npx command. The project has reached ~27.5k GitHub stars, 2.8k forks, and 120k+ npm downloads while running on a fully zero-cost infrastructure: GitHub serves as the CMS and CDN for Markdown/JSON components, Vercel hosts the Astro-based dashboard, Cloudflare Workers handle scheduled doc monitoring and weekly KPI reports, Clerk provides free-tier authentication, and Supabase tracks download analytics. Instead of writing every component, the project aggregates curated community sources (including official Anthropic skills and several popular skill/agent collections) with licenses preserved. Key engineering choices include pre-generated static JSON instead of database queries, deep-merge conflict detection during installation, and graceful degradation on network failures. The post analyzes why the project succeeded—precise timing, zero friction, filling Claude Code's ecosystem gap—and questions its long-term sustainability, quality control, and relationship with potential first-party Anthropic marketplaces.

Introduction

Daniel Ávila, a developer from Chile, built what is effectively a true app store for Claude Code: claude-code-templates (aitmpl.com). Users can browse, search, and one-click install six categories of components—Agents, Commands, MCPs, Settings, Hooks, and Skills—much like installing apps on a phone.

Project stats:

~27.5k stars, 2.8k forks
120k+ npm downloads
Core thesis: configuration is code—and configuration beats code
The most impressive part: total server cost is $0.

Key points: The Zero-Cost Architecture

GitHub as CMS

Agents and Commands live as .md files under cli-tool/components/agents/ and cli-tool/components/commands/; MCPs use JSON configs; Settings and Hooks use JSON plus optional Python/Bash scripts.
The CLI fetches files directly from raw.githubusercontent.com, so GitHub's CDN is the content distribution layer.
Benefits: built-in version control, mature collaboration via PRs, a free global CDN, and zero-friction contributions—anyone who can write Markdown can submit a component.

Vercel-hosted Dashboard

aitmpl.com and app.aitmpl.com run on Vercel with Astro 5, React islands, Tailwind v4, and Clerk for free-tier auth.
Data lives in static files (dashboard/public/components.json, trending-data.json) pre-generated by a Python script that scans the repo—the dashboard never hits a database or the GitHub API, so there's zero latency and no rate limits.
All APIs are Astro API routes; no separate serverless project.

Cloudflare Workers for Scheduled Tasks

docs-monitor: crawls code.claude.com/docs hourly and sends Telegram notifications on changes.
pulse (Weekly KPI Report): every Sunday at 14:00 UTC it collects GitHub stars, Discord activity, Supabase stats, Vercel deployment data, and Google Analytics into a weekly Telegram report. It's a single index.js with zero npm dependencies; every source collector catches its own errors and shows ⚠️ Unavailable instead of crashing.

Supabase for Analytics

CLI installations log records to Supabase—not for real-time recommendations, just aggregated stats powering the dashboard's trending section.

| Function | Solution | Cost | |---|---|---| | Content storage | GitHub repo | Free | | Content delivery | GitHub raw CDN | Free | | Dashboard hosting | Vercel | Free tier | | Authentication | Clerk | Free tier | | Scheduled tasks | Cloudflare Workers | Free tier | | Data tracking | Supabase | Free tier | | Docs site | docs.aitmpl.com | Vercel free |

Total server cost: $0.

CLI Design Philosophy

npx as a service: npx claude-code-templates@latest always pulls the latest version—no global installs, no version fragmentation.
Six component types: Agents (specialist roles like security auditor), Commands (custom slash commands like /generate-tests), MCPs (integrations: GitHub, PostgreSQL, Stripe, AWS), Settings (timeouts, memory, output styles), Hooks (pre-commit validation, post-completion actions), Skills (PDF processing, Excel automation).
Smart installation: deep merges existing .mcp.json instead of overwriting, prompts on conflicts; supports four install scopes (user / project / local / enterprise); adapts python3 to Windows-compatible python/py; batch atomic installs with --agent a --command b --mcp c --yes; graceful degradation on network failures.
Interactive TUI: without arguments, it presents a menu (Analytics Dashboard, Chats Mobile, Agents Dashboard, Project Setup, Health Check). Health Check diagnoses the Claude Code installation before guiding setup.

Community Aggregation Strategy

Rather than writing every component, the project curates existing sources, preserving original licenses and attribution:

anthropics/skills — 21 official skills
K-Dense-AI/claude-scientific-skills — 139 science-domain skills (biology, chemistry, medicine)
obra/superpowers — 14 workflow skills
alirezarezvani/claude-skills — 36 professional-role skills
wshobson/agents — 48 agents
awesome-claude-code — 21 commands

This gave the project 200+ components on day one and an instant community feel.

Data Loop

A trackingService logs command executions, template installations, per-component downloads, installation outcomes (success/failure/duration), dashboard visits, and health checks to Supabase, feeding trending-data.json. A Python script (scripts/generate_components_json.py) scans cli-tool/components/ and regenerates static JSON for the dashboard.

Why 27.5k Stars

1. Precise timing — after Claude Code exploded in early 2025, developers were looking for ways to extend it. 2. Zero friction — no sign-up, no server setup, one npx command. 3. Correct ecosystem niche — Anthropic provides the skills mechanism but no discovery/management layer, much like npm for Node.js or PyPI for Python. 4. Official recognition — the README carries a "Claude for Open Source" badge from Anthropic's official open-source sponsorship program.

Open Questions and Limitations

Sustainability: can free tiers on Vercel, Cloudflare, and Supabase scale if downloads hit millions?
Quality control: curated sources work today, but review burden will grow exponentially with individual submissions; is there automated testing?
Relationship with official efforts: if Anthropic ships a first-party marketplace, what room remains for third-party curation?
Monetization: pure MIT open source today; a "Sponsored by Z.AI" badge is not a business model.

Core Insight

Configuration matters more than code. Daniel didn't train a model—he built a configuration-management layer that makes Claude Code's capabilities discoverable, installable, composable, and trackable. This is the new infrastructure layer of the AI era: not compute, not model weights, but *how model capabilities get organized and reused by humans*. One person, zero servers, 27.5k stars—not because the tech was complex, but because the insight was sharp.