Mechanical fetch — no AI judgment needed. Four metrics, ranks by stars / star-growth, flags AI/dev relevance from topics + description, writes markdown.
The four metrics
- This Week — repos
created:>7d, sorted by stars (REST search API). - This Month — repos
created:>30d, sorted by stars (REST search API). - Fastest Growing (24h) — top AI/dev repos by stars gained today, scraped from
github.com/trending?since=daily. Catches same-day spikes. - Fastest Growing (30d) — top AI/dev repos by stars gained this month, scraped from
github.com/trending?since=monthly. The durable breakouts.
Metrics 3 & 4 are the velocity pair. Unlike 1 & 2 (which filter on
created:<date> and so only ever see brand-new repos), they surface a repo of
any age that is suddenly ripping. A pure stars-sorted API query cannot do
velocity — the top of that list is all mega-repos, so a single-digit-k sleeper
never appears. The trending page is GitHub's own growth ranking.
Output
<VAULT>/inbox/research/github-trending/YYYY-MM-DD-trending.md — VAULT from
AGENTIC_OS_VAULT in ~/.claude/.env, default ~/the-vault.
Sections: Top 10 This Week, Top 5 This Month, Top 10 Fastest Growing (24h),
Top 10 Fastest Growing (30d), Content Radar (AI-pick summary). Each repo block:
stars, language, created date, topics, description, [AI/DEV] flag. If a
scrape fails, that section degrades gracefully and the others still render.
Manual invocation
python ~/.claude/skills/github-trending/scripts/fetch.py
AI/Dev classifier
Substring match (case-insensitive) on topics OR description against:
ai, llm, gpt, claude, anthropic, agent, mcp, prompt, embedding, rag, deepseek, kimi, gemini, openai, copilot, codex, fine-tun, transformer, ollama, langchain, chatgpt
Not perfect — false positives on "DarkGPT"-style spam, false negatives on AI projects with vague descriptions. Good enough for a glance.
Rate limit
Per run: 2 Search API calls + 2 trending-page GETs + up to ~1 enrich call per
trending repo scanned (GET /repos/{full}). Unauthenticated Search = 10
req/min; core REST = 60/hr per IP — fine for a few runs/day, but bursty manual
re-runs can exhaust it. Set GITHUB_TOKEN in ~/.claude/.env to lift both to
5000/hr.