Big ideas in small bites. An Instagram-style feed of AI-written idea cards: pick a topic, scroll, double-tap to like, bookmark to save, keep a daily streak. No account, no backend to run — it's a single static page plus one serverless function, deployed on Cloudflare's free tier.
public/index.html— the entire frontend, no build step, no framework.functions/api/ideas.js— a Cloudflare Pages Function that calls Workers AI (Llama 3.3 70B, with Qwen3 and Llama 4 Scout as fallbacks) and returns 5 idea cards for a topic.public/manifest.json,sw.js,icons/— installs as a real PWA; add it to your iPhone's home screen and it runs full-screen.
No API key anywhere: idea generation runs on Cloudflare Workers AI, which includes 10,000 free neurons a day (resets at 00:00 UTC).
git clone https://github.com/RouterGlock/an-idea.git
cd an-idea
./deploy.shThe script logs you into Cloudflare, creates the an-idea Pages project,
and deploys. Then on your phone: open the *.pages.dev URL in Safari →
Share → Add to Home Screen.
.github/workflows/deploy.yml redeploys on every push to main (and on
manual "Run workflow"). It needs one repo secret — a Cloudflare API token
with the Account · Cloudflare Pages · Edit permission:
gh secret set CLOUDFLARE_API_TOKEN --repo RouterGlock/an-ideaThe account ID is inlined in the workflow (it isn't secret). Once this is
set, deploy with git push instead of ./deploy.sh to avoid mixed
deployment sources.
npx wrangler pages dev public --remote # --remote needed: AI models don't run locallyFree. Each topic load is one Workers AI call (~1.5k tokens) — roughly 30–60
loads a day inside the free 10k-neuron pool on Llama 3.3 70B; swap the first
entry in MODELS for a smaller model if you want more headroom. If the daily
pool runs out, calls fail until 00:00 UTC and the app shows a retry button.
Cloudflare retires models occasionally — run npx wrangler ai models list
to see what's current.
MIT © 2026 RouterGlock — see LICENSE.