Sooo, I'm always looking for new ways to improve how I use AI agents and build better workflows around them. And there's A LOT of GitHub repos out there. Some are genuinely useful, some aren't, and it's easy to get lost jumping from one repo to another.
So I built a small page to make it easier for me to find interesting projects and keep track of what's new.
Why not just use GitHub's trending page?#
GitHub's trending page shows the top 25 repos across everything. Niche AI agent tools (skills, MCP servers, memory, token savers) rarely make that list, and directories don't know what I care about. So the page only watches my niches, and it updates itself every day.
What it watches#
GitHub Trends collects trending repositories across four areas I'm particularly interested in: skills, MCP servers, memory and context tools (like MemPalace), and agent tooling that helps reduce token usage (like Caveman). Each one is a tab on the page.
| Tab | What goes in it | Example |
|---|---|---|
| Skills | Claude and agent skills | anthropics/skills |
| MCP | Model Context Protocol servers | upstash/context7 |
| Memory & Context | Long-term memory for agents | MemPalace/mempalace |
| Agent tooling | Token savers, gateways, code intelligence | JuliusBrussee/caveman |
How it works#
There's a small script that runs every morning through GitHub Actions. It searches GitHub, filters out the noise, tracks star gains, and marks repositories I haven't seen before as NEW.
flowchart LR
A[GitHub Action<br/>06:00 UTC daily] --> B[fetch.mjs]
B --> C[GitHub search API<br/>25 topic and keyword queries]
B --> D[github.com/trending<br/>daily and weekly]
B --> E[Watchlist<br/>repos I never want to miss]
C & D & E --> F[Filter noise<br/>keywords, exclude list, max 2 tabs per repo]
F --> G[Rank and cap<br/>30 per tab]
G --> H[(repos.json<br/>seen.json<br/>history.json)]
H --> I[Next.js page<br/>4 tabs, search, NEW marks]The three data files each have one job:
| File | Job |
|---|---|
repos.json | What the page shows |
seen.json | Every repo already shown, so new ones get a NEW mark |
history.json | Daily star counts, so the page can show stars gained this week |
Kept small on purpose#
The page itself is intentionally simple: monochrome design, light and dark themes, search, and a copy button for installation commands.
- No database, no server, no accounts. One script, three JSON files, one page.
- No dependencies in the fetch script.
- About 2.5 minutes per daily run, free on GitHub Actions.
How I built it#
I built it using Claude Code. Opus handled planning and reviewing the work, while Sonnet agents worked on building the different pieces.
Nothing too complicated. Just a small tool that updates itself and helps me spend less time searching and more time actually exploring interesting projects.
The code is open source: gaderOussama/github_trends.
