Cloudflare launches Worker Previews: isolated preview environments for every Git branch

What Cloudflare shipped

Cloudflare has launched Worker Previews, announced on its blog on September 22, 2026. The feature gives every Git branch its own production-like place to run, carrying separate code, configuration, URL, observability data and state. According to the Cloudflare blog, a preview is created with a single command, npx wrangler preview, and every push to the branch updates the same running preview, so a stable URL can be shared and exercised with requests, clicked through manually, or probed from CI while iterating.

Cloudflare frames the launch against a familiar pain point: changes that behave one way in staging and another in production. Previews are intended to sit close enough to production that runtime behaviour — an API endpoint returning the right response, or a UI change judged in context — can be validated before a merge.

Stateful resources are isolated per branch

The most technical part of the design concerns Durable Objects. These run on a singleton model, where one instance owns a given object ID and its storage. Cloudflare points out that a preview sharing production's namespace would not just read stale data — it could modify the same instance serving live traffic. To prevent this, the company automatically provisions a new Durable Object namespace and Container application for each preview, so a failed migration or a bad schema change stays contained within that branch.

In code, these resources are reached through ctx.exports: the same expression resolves to the production namespace in production and to the preview's own namespace inside a preview. Cloudflare suggests this opens the door to parallel experiments, such as running different configurations on separate branches to compare cold-start and warm-start performance side by side.

One Worker, many environments

In the Cloudflare dashboard, previews work like branch switching: a breadcrumb next to the Worker's name, which defaults to Production, lists every preview alongside it. Unlike Wrangler environments, which require deploying and managing a separate Worker for each environment, previews keep all of that isolation within a single view. Cloudflare says hundreds of previews can run at the same time without interfering with one another or with production.

A base configuration plus per-preview overrides

Each preview starts from a copy of a base configuration — variables, secrets, bindings and settings that the developer defines — mirroring how a branch forks from main. An individual preview can override that configuration when it needs to, for instance pointing at its own database or a test API key, without touching production, the base, or other previews. Preview URLs can also be served on a custom domain, which Cloudflare notes matters for anything hostname-sensitive: authentication providers, cookies, CORS and OAuth redirects then behave the way they will in production.

Observability, and a role for agents

Workers Observability applies per preview. Cloudflare says each request is traced across its whole lifecycle — fetch calls, binding operations and handler invocations — so a failure can be diagnosed without sifting through production traffic or signals from other changes.

The launch is explicitly aimed at AI-assisted development. Cloudflare describes an "Agent Development Lifecycle" in which each change stands on its own, can ship independently, can be observed, and can be revised before it reaches production. Agents can send traffic to a preview URL, and with Browser Run they can open it in a headless browser, step through a login flow, capture screenshots or record the session as replayable DOM events, and connect a failed request to observability events from the same run.

Why it matters

Shared staging environments are a long-standing bottleneck: contributors queue for one environment, and drift between staging and production makes results unreliable. Per-branch previews make pre-production testing parallel and disposable, and the automatic isolation of Durable Objects and Containers extends that safety to stateful code — precisely the area where naive preview setups are most dangerous. The agent framing is also significant: as agents push more and larger changes, verification becomes the constraint, and Cloudflare is betting that giving every change its own observable, throwaway environment is how verification scales. One caveat worth noting: all of these claims come from Cloudflare's own announcement, so practical details such as limits, pricing and real-world performance will only become clear through hands-on use.