Cloudflare OS – an open-source AI productivity environment

Cloudflare OS is positioned as an internal operating system for AI productivity, now made available under an open-source license so other companies can adopt or adapt it. It replaces the traditional SaaS model by letting each user spin up a private instance of an AI-generated app, called a gadget, which runs in isolation. This shift changes how engineers think about deploying AI-augmented tools, moving from shared services to per-user sandboxed workloads.

The platform’s three core pieces are a chat UI for issuing commands to AI agents, a sandbox that builds and runs gadgets, and Gatekeepers that act as capability-based guards for any external service the gadget contacts. Gatekeepers wrap external APIs, handle OAuth, restrict access to the exact resource requested, and log every operation, providing a human-in-the-loop checkpoint for side-effecting actions. For engineers, this means that AI-driven automation can be granted narrow permissions without exposing broader system privileges.

To try the system, developers install pnpm, run a local script, and access a web UI on localhost; the stack runs on Cloudflare’s Wrangler and workerd runtimes. This local mode is explicitly marked as unsuitable for production, and deploying to a real Cloudflare account requires additional steps outlined in the repository. Adoption therefore incurs the cost of setting up the development environment and potentially integrating required services such as GitHub or Google Docs, which need separate configuration.

The early-access label signals that many features are still rough and that the platform may not handle high-scale production workloads reliably. External integrations only work after explicit setup, and the human-in-the-loop approval model can introduce latency when agents request privileged actions. Engineers must therefore evaluate whether the sandboxed approach fits their reliability and performance requirements before moving beyond experimentation.

If an organization adopts Cloudflare OS, engineers can quickly prototype AI-generated utilities, like slide decks or issue dashboards, without writing code from scratch, while security teams retain visibility through Gatekeeper logs. However, the need to maintain the sandbox infrastructure and manage the early-access limitations means that operational overhead and risk assessment remain necessary. The open-source nature allows teams to customize the OS to their own policies, but they must also assume responsibility for hardening the deployment for production use.