Three sources, one shape
Registry, Glama and GitHub all describe a server differently. A normalisation pass maps them onto a single record and merges duplicates by repo URL, so the same server never shows up twice under two names.
All work
A self-updating catalogue of the AI tooling ecosystem
AI Library
- Role
- Creator · full stack
- Year
- 2026
- Impact
- Refreshed daily, fully static
Overview
Finding a good MCP server or Claude Skill means trawling registries, GitHub topics and half-maintained gist lists. AI Library collapses that into one searchable directory that refreshes itself every day.
Every entry carries setup instructions generated per assistant — Claude, Cursor, Codex, Windsurf, Cline and VS Code each get the exact config block they need, so discovery and installation are one step instead of two.
What I built
- Designed the whole product end to end — data model, ingestion, UI and deploy.
- Built the ingestion layer over the MCP Registry, Glama and the GitHub Search API, with normalisation and de-duplication across sources.
- Wrote the per-assistant snippet generator that turns one canonical server record into six different install formats.
- Ran the entire catalogue as build-time static data — no database, no runtime API, nothing to page me at 2am.
The hard parts
Fresh without a server
A daily GitHub Actions cron re-runs ingestion and commits the regenerated dataset, which triggers a rebuild. The site stays fully static while the content behaves like it is live.
Rate limits
The GitHub Search API is tight. Requests are batched, cached between runs and backed off on 403s, so a full refresh finishes inside one free-tier Actions window.
Outcomes
- Assistants supported
- 6
- Auto-refresh
- Daily
- Servers to maintain
- 0
Stack
- Next.js 16
- TypeScript
- Tailwind v4
- Motion
- GitHub Actions
AI in the build
Generates per-assistant install and usage snippets from the MCP Registry, Glama and the GitHub Search API