Cloudflare OS

Cloudflare OS is an open source AI productivity environment from Cloudflare that turns agent work into private, shareable workspace objects called gadgets. It is not a desktop operating system. The README uses the OS metaphor because the product manages users, workspaces, applications, permissions, external services, and AI agents on top of Cloudflare Workers. The project was originally built for internal Cloudflare teams and is now available for others to copy, run, and adapt as a company-specific AI workspace. The product centers on an agent chat UI where people can ask agents to create documents, build small apps, prepare slide decks, repair text in connected documents, or make issue dashboards from attached repositories. Instead of sending everything into one shared SaaS app, Cloudflare OS creates a private instance of each generated app for the user. Cloudflare calls these apps gadgets. A gadget can be shared later, but the default model is private and sandboxed. That design matters because agent-built apps often handle sensitive data and code written on the fly. The second major piece is the Gatekeepers security framework. Gatekeepers sit between agents, gadgets, and outside systems such as CRMs, Google Docs, GitHub, or other business tools. The README describes Gatekeepers as capability-based services that wrap external APIs, handle authorization, enforce narrow access, log actions, and support human review for side-effecting work. A notable detail is asynchronous approval: the Gatekeeper can simulate a pending action so the agent can continue, then let the human approve or reject actions afterward. Cloudflare OS is built on Workers, Durable Objects, Dynamic Workers, and Facets. Each workspace maps to its own Durable Object, and gadgets run in Dynamic Worker facets. The project can be run locally with `pnpm run-local`, which starts the stack on wrangler and workerd, or deployed to a Cloudflare account through the linked deployment flow. The README warns that the August 2026 v2 release is early access and still rough, so teams should expect active development rather than a polished hosted SaaS. For builders, Cloudflare OS is most interesting as a reference architecture for safe agent workspaces. It shows how to combine generated apps, collaboration, identity, company context, per-user sandboxing, approval queues, and external service connectors in a Workers-native stack. Pricing depends on where you run it. The code is open source, but production use will incur Cloudflare, model, storage, and integration costs based on traffic and workload size. The repo is also useful even if a team never deploys it unchanged. The architecture shows how Cloudflare thinks about AI agents as accountable workers with constrained access, private execution spaces, shared artifacts, and reviewable side effects. That makes it a strong reference for teams building their own internal agent platforms on Workers or similar edge runtimes.