Cloudflare open-sources vibe-coding platform for people who aren't coders

Cloudflare spent months building an internal tool called Cloudflare OS so employees who aren't engineers could describe a workflow in plain language and have an AI agent turn it into a working app. The company says thousands of staff now use it daily to whip up documents, slides, small data-visualization tools, and other one-off apps. Now Cloudflare has put the whole thing on GitHub for anyone to run.

The pitch isn't just "AI writes your app." It's that the sandboxing is tight enough to let non-technical people vibe-code without creating a security disaster. Principal engineer Kenton Varda described it as a setup secure enough that security teams could let untrained users loose and still sleep at night. Under the hood, each app instance runs in its own isolated environment, built on a feature called Dynamic Workers. Instead of standard containers, it uses lightweight V8 isolates that start in milliseconds and use only a few megabytes of memory — Cloudflare claims that's up to 100 times faster and dramatically more memory-efficient than containers. AI agents start with zero permissions and have to explicitly request access to anything, and both server and client code are cut off from the open internet unless a capability is granted.

CIO Sam Rhea's blog post gets into the messier lessons. Cloudflare initially handed non-engineers the same coding harnesses engineers use, just with friendlier interfaces, and the result was a pile of vibe-coded apps nobody actually needed — solutions hunting for problems. The company also realized that letting AI agents write code faster meant more bad code got written faster too, so it built something called the Cloudflare Engineering Codex to give both humans and AI agents a standard to check work against. Over four months, the AI reviewer built on that codex flagged nearly a quarter million deviations from engineering standards and blocked 16,000 merges, according to systems engineer Timo Reimann.

The platform also lets admins cap AI spending and set rate limits per user, a nod to how easily teams burn through model-usage budgets — Rhea specifically pushed back on the idea that every employee needs the priciest frontier model just to summarize an inbox. Whether that restraint holds once companies start pressuring staff to use more AI regardless of need is a separate question the article doesn't pretend to answer.

The open-source release came with a rough edge: running the Cloudflare OS backend requires a paid Workers plan, and that wasn't obvious upfront. A GitHub user on the free tier got partway through deployment before being blocked, and complained about wasting time on a requirement that should've been flagged immediately. Cloudflare fixed the messaging quickly and responded to the complaint, but it's a reminder that "open source" doesn't always mean frictionless, or free.