Cloudflare CIO: What We Did When Employees Started Vibe Coding Apps

TLDR

• Cloudflare built CloudflareOS after employees began vibe coding applications but needed a secure way to connect them to company systems and deploy them.
• The internet infrastructure company looked at third-party tools but they all had limitations and didn’t meet Cloudflare’s business needs. So it decided to build its own tool for internal use – and then open sourced it.
• CloudflareOS has an unusual feature called Gatekeepers that can simulate the outcome of an AI agent’s proposed action, letting the agent continue working while the real task waits for human approval.

Cloudflare CIO Sam Rhea knew his company faced an AI-related dilemma when an employee approached him with a promising new app − and a rather unusual request.

A member of its go-to-market team had spent a weekend “vibe coding” what he described as a “super app” that could transform the way the company sold its products. In vibe coding, a user creates an app by describing it in plain language, and AI writes the code to build it.

But there was a catch: He needed supporting infrastructure to make the app work: about a dozen API keys to production systems, help deploying the application without accidentally exposing it to the public internet and instructions for setting up a continuous integration and deployment pipeline.

The internet infrastructure company faced a dilemma: Generative AI gave employees who didn’t know how to code the ability to build useful applications and automations. Forward-thinking companies would not want to suppress these acts of innovation. But these activities also bring security, governance and deployment concerns.

At first, Cloudflare looked at third-party tools, but saw drawbacks. “They all have their own limitations,” Rhea told The AI Innovator. “We would love to have used many of those tools out there. It would have meant that months ago we didn’t have to go on this journey of building this out necessarily. But they have their limitations, which is what really kicked us down this road.”

The downsides were major: Third-party tools often lock users into a vendor, offer only a niche capability rather than an all-purpose platform, or just provide some aspect of security and governance guardrails, Rhea said. Cloudflare wanted a platform that had all of the above and more.

“So we did what we kind of almost always do at Cloudflare,” he added. “We said, ‘well, let’s use our own primitives and our own components and build something and see how that goes.’”

Making it open source

That’s when Cloudflare built CloudflareOS to meet the company’s needs – and open-sourced it on GitHub for others to freely download, modify and deploy under an Apache 2.0 license, including for commercial uses. Users can run the entire stack locally for testing or deploy it to a Cloudflare account, which has a free tier.

Accessible from a browser, CloudflareOS lets apps connect only to approved databases and systems, lets users set rules on access, and applies a zero-trust level of security, meaning no user or device is automatically trusted.

Cloudflare says a large portion of its workforce, spanning engineering, sales and other functions, now uses the platform daily.

Employees began finding their own uses for it. Cloudflare interns were sent around the company asking teams what work they hated and then using CloudflareOS to automate it. Salespeople used it for territory planning and competitive intelligence. Other workers created applications without first having to learn command-line tools, Git repositories or conventional application deployment.

Then Cloudflare’s customers started asking for the same thing. Rhea said customers were attracted by the ability to choose among AI models rather than being locked into a single provider, run AI workloads in cloud infrastructure instead of on individual devices, and place security controls between AI agents and enterprise systems.

“Teams want optionality from a capabilities perspective, but they also want optionality from a cost perspective,” Rhea said.

CloudflareOS has controls for setting token and cost limits. It can route requests to different AI models depending on the task, meaning that the most advanced and expensive models could be used for the most complex tasks while cheaper models could be used for simpler activities.

From tax IDs to sales territories

AI can be used in different ways at Cloudflare. Employees even have a dedicated email where they send their ideas: AIwins@cloudflare.com.

“It’s really everything from people … automating more deterministic workflows in a safe place where they don’t have to go learn how to deploy an application to taking our organization’s brain and helping people be just so much more productive, so they can focus on spending more time with customers or building better technology,” Rhea said.

In one case, an intern worked with Cloudflare’s European tax organization on the tedious process of reconciling tax identification numbers stored in different country formats. Instead of repeatedly feeding spreadsheets to a large language model and asking it to perform the task, the intern used AI to create a deterministic application into which employees could paste files.

“AI is a much better tool maker than a tool,” Rhea said.

Sales territory planning exemplifies another role for AI. Cloudflare created capabilities and workflows combining its territory-planning methods with data from enterprise systems. Rhea said a salesperson in London who had been on the job for three weeks created a detailed territory plan in minutes − work the employee said would otherwise have taken days or weeks.

Competitive intelligence provides another example. Cloudflare has encoded information about competitors by industry, sector and product line into organizational context that salespeople can use to generate customized battle cards ahead of customer meetings.

“It’s making prescriptive recommendations such that a sales team can review it together, and we’re all working from some of the same source material and recommendations and guidance,” Rhea said.

“And then if we want to make changes, it allows us to focus on places where there might be some ambiguity where we really want human judgment, like that customer in this industry has this really specific challenge, and this member of the team has this background and this experience on that particular type of challenge. Let’s match the two,” he continued. “That’s the kind of human judgment that you’d want to use.”

An approach with an unusual feature

CloudflareOS combines an agent chat interface with company knowledge, AI-generated applications and a security architecture governing how those applications and agents interact with outside data and systems.

Its most unusual feature may be what Cloudflare calls “Gadgets.”

Rather than sending a user to a centralized software program, CloudflareOS can create a private, sandboxed instance of an application for that individual. A worker might ask an agent to build a slide presentation, dashboard or other small application. The resulting Gadget can then be modified through further AI prompts and securely shared with colleagues.

For decades, SaaS vendors have built centralized applications and chosen the features users get. With generative AI, employees can increasingly ask agents to create or modify the software features they actually need.

CloudflareOS secures the task through a framework called Gatekeepers.

Agents and Gadgets don’t automatically receive blanket access to an employee’s connected enterprise resources. When a user introduces one to an outside resource, a Gatekeeper mediates the connection, handles authorization, limits access to the intended resource and logs actions.

For operations with side effects, it can also require human approval. Cloudflare has devised an unusual approach to keep those approvals from constantly stopping an autonomous agent. A Gatekeeper can simulate the outcome of a proposed action so the agent can continue working, while the real action waits for people to approve or reject it later. Meanwhile, its broader security stack can add controls around enterpris