Cloudflare Open-Sources Secure AI Platform for Non-Technical Staff to Build Custom Apps

Cloudflare has made Cloudflare OS publicly available on GitHub, releasing a platform designed to let enterprise teams build customized applications powered by AI while maintaining strict security boundaries. Unlike typical cloud software where users interact with a shared service, Cloudflare OS grants each worker their own isolated copy of any application they create, letting them modify it on the fly through natural language prompts.

A New Model for Enterprise Software#

Rather than the standard multi-tenant SaaS approach, Cloudflare OS instantiates dedicated, sandboxed instances of applications for individual users. This architectural shift—powered by Cloudflare's open-source workerd runtime and Dynamic Workers—enables a fundamentally different security posture. Lead architect Kenton Varda framed the distinction as bridging consumer AI flexibility with enterprise governance: "a chatbot with connectors, just like every other tech company is doing. […] Except actually, it's different." Users can now ask AI agents to add features directly to their personal app instances without waiting for centralized software updates.

How Isolation Prevents Data Leaks#

The platform's security foundation rests on Gatekeepers, a capability-based access control system that strictly limits what resources each isolated app instance can reach. Unlike standard Model Context Protocol (MCP) connections that often grant broad ambient permissions, Gatekeepers enforce zero-trust principles: agents begin with no permissions and must be explicitly granted access to specific resources. Sensitive database columns remain masked, role-based rate limits apply, and destructive operations require human approval. When someone shares an app (called a "Gadget" internally), the system verifies that the recipient already has permission to access all connected resources—no code bug can accidentally escalate privileges.

Internal Results Driving Open Release#

The platform emerged from Cloudflare's own operational bottlenecks. Employees had been deploying unvetted AI workflows to automate repetitive tasks, but doing so safely required careful access control. A human-staffed "magic AI email" alias initially triaged requests, revealing that off-the-shelf agent frameworks handled routine scripting well but struggled with context-dependent workflows. Since May 2026, Cloudflare staff have built over 4,000 custom business tools in 30 days; sales teams recovered roughly 10,000 hours of manual work on territory planning. Engineering teams deployed automated review agents that flagged nearly 250,000 potential bugs and blocked 16,000 non-compliant pull requests against a machine-readable policy codex.