Start here
A Cloudflare Worker intercepts every request before it reaches your origin. It can run AI models, cache in R2 or KV, rewrite HTML, detect the visitor's country — all at the edge, with 0ms cold starts (V8 isolates eliminate the VM boot model), in 330+ cities. Your origin receives a completely normal request and never knows any of this happened.
AI · $0 egress · semantic search · SEO · voice · video · caching · geo pricing · subtitles · collab room designer — none touched the origin.
The concept every other demo here relies on: a Worker sits in front of your site via a route pattern, transforms the response, and your origin never knows it exists. ~8 lines of code, zero backend changes.
Learn more →HTMLRewriter injects a live chatbot into any existing page, grounded in your real product catalog via AI Search — not generic LLM knowledge. Zero origin changes, one Worker route.
Learn more →Watch the AWS/GCS egress cost ticker climb in real time as images load, while R2 stays at $0.00 for the exact same traffic. Swap the storage binding, keep everything else.
Learn more →"Comfortable reading chair" finds the POÄNG Armchair even though not one word matches. One embedding call + a Vectorize query — ~30 lines, no Elasticsearch cluster to run or maintain.
Learn more →A Worker enriches your existing reviews API with real-time sentiment scores — the backend behind it never changes. distilbert costs $0.026/M input tokens; scoring 1M reviews ≈ $3.90.
Learn more →5 lines read request.cf.country, 3 more convert the price. ฿ / Rp / S$ / RM shown to every visitor — zero geo-IP API subscription, zero added latency.
Learn more →A ~10-line cache-aside pattern turns 10,000 origin calls/hour into 166 with a 60-second TTL. Hot reads land in 500µs–10ms from the nearest Cloudflare PoP.
Learn more →Upload a clip — Workers AI routes to the right STT model per language, runs a two-pass AI correction, and exports a timestamped .vtt. One POST request; AWS needs S3, IAM, and a batch pipeline.
Learn more →Speech-to-text, LLM reasoning, and text-to-speech chained in one Worker endpoint — no separate voice infrastructure to provision, reply back in under 2 seconds.
Learn more →No Worker code at all — just a URL prefix. Converts to WebP/AVIF and resizes per device automatically: 1.1 MB shrinks to 15 KB on mobile, measured live on this page.
Learn more →Four AI-powered endpoints your origin has never seen — style match, assembly guide, mood discovery, watchlist planner. ~30 lines each, deployed in 15 minutes, zero backend sprint.
Learn more →Marketing uploads a video — Stream auto-encodes every quality level, delivers adaptive bitrate from 330+ cities, and builds the player. Zero engineering tickets, $0 egress.
Learn more →A blank textarea kills review completion. Workers AI generates tone-matched tags and a full draft review in under 2 seconds — completion rate roughly triples.
Learn more →Your React SPA sends Googlebot the same blank title on every URL. A Worker injects unique titles, meta descriptions, and JSON-LD via HTMLRewriter — no SSR migration required.
Learn more →What gets scoped as a 6-month build — WebSocket server, Redis pub/sub, load balancer — is one ~400-line Durable Object class here. SQLite persists state; Hibernation makes idle connections free.
Learn more →Point your phone camera — Workers AI finds objects at the edge in under a second. Fixed 80-class detection for a fraction of a cent, or open-vocabulary/caption/OCR via Moondream on the same endpoint.
Learn more →Ask any question about a photo in plain English — "does this look damaged?" — with multi-turn follow-ups. One natively multimodal model, no separate vision pipeline to build.
Learn more →Upload a photo and describe an edit in plain English, or skip the upload and generate one from scratch. One Workers AI model, two modes, no diffusion pipeline or masking UI to build.
Learn more →Object counting, baby distress, fall detection, exam-integrity flagging, safety flagging — five monitoring ideas, one Workers AI model, five prompts. Proves how cheaply a new AI use case can be prototyped before committing engineering budget.
Learn more →A classic FAQ search box — but retrieval is one AI Search call, not a Vectorize index you build and own. Cites the exact source document behind every answer, and honestly says "I don't know" when nothing relevant is found.
Learn more →A real checkout — cart, email, shipping address — running seven durable steps: validate against D1, charge, invoice, render a real PDF via Browser Rendering, sleep for the warehouse, email that PDF, notify fulfillment. Watch the payment step fail and retry with backoff live, then download the actual invoice PDF.
Learn more →For your engineering team
Two production-ready Workers — AI chatbot and R2 image cache. No ORIGIN constant to configure, no origin to touch. Deploy to any hostname, bind a bucket or an AI binding, done. Your team can have either live in under an hour.