Four free Cloudflare features, live

1 · Durable Objects: an atomic counter

A strongly-consistent, single-instance stateful object. The homepage's visitor counter uses KV, which is eventually-consistent and capped at 1,000 writes/day, so a busy day silently stalls it. A DO increments exactly once, no race, 100k writes/day. It used to cost money; it's free now, and it fixes that counter cleanly.

env.COUNTER.get(id).fetch()  // SQLite-backed, atomic put

2 · Workers AI: caption a photo

Every model in the catalog, 10,000 neurons/day free. This runs @cf/llava-1.5-7b over a real grid photo to generate alt-text. A small, fast vision model hallucinates charmingly now and then (it once called a cloudy sky a UFO). One binding call does it: no GPU, no key.

env.AI.run("@cf/llava-1.5-7b", { image, prompt }, { gateway: { id } })

That third argument arrived in Aug 2026, when Cloudflare folded Workers AI and AI Gateway into one control plane. Before that, the gateway was a separate proxy you pointed a client at. Now it's an option on the call you were already making, and naming one gets the account full request and response payloads, token counts per model, and cost attribution, with nothing to set up first. Section 3's log line knows this call took about five seconds. It has never known what it cost.

Caching is off here on purpose. The button sends a byte-identical request every time, so one cached answer would quietly become every visitor's answer and the neuron counter would stop moving. The lede promises nothing is faked, and a replay is faked.

3 · Workers Logs: structured observability

200,000 events/day, 3-day retention, free. Every button above emitted a structured log line from the worker ({ feature, ms, … }), filterable and searchable in the dashboard. I can't show it here (the logs live in Cloudflare, not the page), but it's on: [observability] enabled = true in wrangler.toml. Click around, then watch them stream in Workers & Pages → cf-garage → Logs.

console.log(JSON.stringify({ path, feature, ms }))

4 · Custom spans: trace a request

Workers auto-trace every platform call (each fetch, KV read, DO call) into a waterfall in Observability. As of Jun 16 2026 you can wrap your own logic in spans with ctx.tracing.enterSpan(), nested correctly. This endpoint runs a three-step pipeline (a Durable Object peek, a subrequest, and a compute loop), each in its own span with attributes. The canonical trace lands in Observability; the waterfall below is reconstructed live from the same span timings. (Run it twice: a cold Durable Object shows up as a fat first bar, then the warm path collapses to a few ms.)

ctx.tracing.enterSpan("do.counter.peek", async (span) => { span.setAttribute("counter.value", n) })