GitTok — An open-source "TikTok" for GitHub, in Chinese

English | 中文

GitTok automatically aggregates the best projects from GitHub and the global AI ecosystem every day — with LLM-powered Chinese summaries, classification, and scoring. Browsing GitHub feels like swiping short videos: fun, addictive, and always something new.

✨ Features

• 🃏 Addictive card feed: card-based browsing with personalized recommendations — the thrill of short videos, the density of a real feed. Like / dislike / bookmark
• 🧠 AI Chinese recommendations: LLM scoring (Zhipu GLM-4.7-Flash, free) writes a Chinese summary for every project
• 🗂️ Smart sections: Dynamic channels (Recommended / Hot / Daily / Fun / Following) + Zones (AI / Resources / Tools / Creative)
• 👤 Personalization: your likes and dislikes train a personal feed (localStorage, no account needed)
• ⭐ Following: follow creators you like — their projects and in-library repos they starred land in your Following channel (stored in your browser, no account needed)
• 🔍 Search: full-library search with weighted ranking
• 📱 Mobile-ready: bottom navigation + channel drawer on small screens
• 🌍 Multi-source: GitHub Trending + Search, Hacker News, Product Hunt, ArXiv, Hugging Face, Dev.to, Lobsters, Anthropic/OpenAI

🚀 Live demo

🌐 https://chestnuts-sisyphus.github.io/gittok/

No login required. Data updates automatically every day at 04:00 CST.
The live feed holds 2,831 cards (measured 2026-09-22); 950 of them pass the in-site copy gate into the recommendation pool.

The feed: dynamic channels (Recommended / Hot / Daily / Fun / Following) on the left, zones
(AI / Resources / Tools / Creative) below, heat-ranked Chinese cards on the right.

The agent page: four anonymous read-only access
paths (Agent Skill / MCP server / RSS / REST API) plus a live availability self-check.

📣 Push channels

Reports also leave the website: every digest run pushes to Telegram and Feishu (plus the RSS feed above).

Telegram (@agents_radar)	Feishu group

Both channels are inherited from the upstream project (see lineage at the bottom of this page), which is
why the posts are signed agents-radar.

🏗️ Architecture

GitHub Actions (daily-digest.yml) — daily 04:00 CST
  ├─ Fetch 10+ sources in parallel (GitHub / HN / PH / ArXiv / HF / Dev.to / ...)
  ├─ LLM scoring in Chinese (Zhipu GLM-4.7-Flash, free)
  ├─ Generate digests/ daily reports + data/feed.json (feed data)
  ├─ Auto commit + deploy to GitHub Pages
  └─ Push to Telegram / Feishu / RSS

📦 Local development

# 1. Install dependencies
pnpm install

# 2. Configure environment (copy .env.example to .env)
cp .env.example .env
# Fill in ZHIPU_API_KEY (free tier: https://open.bigmodel.cn — Zhipu AI console → API Keys)
# Fill in GITHUB_TOKEN (optional, raises API rate limits)

# 3. Run the full pipeline
LLM_PROVIDER=zhipu pnpm start

# 4. Frontend
cd web && pnpm install && pnpm dev

🔑 Environment variables

📄 Data sources

🤖 Agent interface (anonymous, no API key)

GitTok speaks to agents as well as humans — every endpoint below is public, read-only, no key, no login.
The four access paths (Agent Skill / MCP server / RSS / REST API) are laid out, with a live
availability self-check, on the Agent page.

# today's arrivals, ranked by the site's own heat score
curl -s https://chestnuts-sisyphus.github.io/gittok/data/feed.json -o feed.json
python -c "
import json,datetime
c=json.load(open('feed.json',encoding='utf-8'))
t=datetime.datetime.now(datetime.timezone.utc).date().isoformat()
f=[x for x in c if (x.get('pushedAt') or '')[:10]==t]
f.sort(key=lambda x:x.get('heatScore') or 0,reverse=True)
print(len(f),'new today; top:',[x['repo'] for x in f[:5]])"

MCP server — tools search / top / detail, zero-install single file, config for Claude Desktop /
Claude Code / Cursor: see mcp-gittok/README.md.

curl -fsSL -o ~/gittok-mcp.mjs \
  https://cdn.jsdelivr.net/gh/Chestnuts-Sisyphus/gittok@master/mcp-gittok/dist/index.js
node ~/gittok-mcp.mjs --selftest

Agent Skill — "read today's GitTok highlights and explain them" (skills/gittok/, Agent Skills format):

node skills/gittok/scripts/hotspots.mjs --limit 12    # from a clone

Full field reference, ranking semantics and etiquette: docs/API.md.

Loading behavior (Service Worker & prefetch) — data endpoints unaffected

The site pre-caches its shell via a build-generated Service Worker (sw.js): fingerprinted
assets/*, fonts/*, index.html, favicon.svg only. All data/*, digests/*, feed.xml,
manifest.json requests pass straight through to the network — the SW never serves a second,
stale copy of the feed. In the browser, the app also prefetches detail shards for currently
visible cards during idle time ("visible = fetchable"), which is site behavior only and does not
change the shard API. Same-day IndexedDB caching is app-layer and expires naturally. See
docs/API.md § 站点加载行为.

🎯 What does "Tok" mean?

GitTok is not a fullscreen clone of TikTok. "Tok" is the feeling: browsing GitHub projects should be as fun and addictive as scrolling short videos. GitTok delivers that with an algorithm-driven card feed — personalized recommendations, Chinese summaries for every project, and fresh picks every day. Swipe through GitHub like short videos, with the density of a real feed.

📜 Tech stack

• Backend: Node.js + TypeScript, driven by GitHub Actions cron
• Frontend: React + Vite, deployed on GitHub Pages
• LLM: Zhipu GLM-4.7-Flash (OpenAI-compatible endpoint, permanently free)

🪜 Lineage

GitTok grew out of agents-radar, a daily digest site for
AI CLI / agent tooling. Its interface below is what the current card feed replaced — the pipeline
(multi-source fetching → LLM scoring → daily digest → push) is the part that carried over.

Upstream site (Chinese)	Upstream site (English)

📝 License

MIT License

Forked from agents-radar, reimagined as GitTok by Chestnuts-Sisyphus.