Cloudflare launches open-source Cloudflare OS

Cloudflare has introduced Cloudflare OS, an open-source, browser-based platform designed to connect artificial intelligence agents, corporate data, and internal workflows within a secure workspace. Built on Cloudflare Workers, Dynamic Workers, Durable Objects, and the company's Access zero-trust network tool, this environment runs directly inside a customer's Cloudflare account. It allows users to interact with agents via a conversational interface to build full-stack applications, generate documents, and automate tasks without using a terminal. Because it is open-source and model-agnostic, organizations can integrate it with various AI models from providers like OpenAI, Anthropic, Google, Microsoft, or Meta.

To manage agent permissions, the platform employs governed connectors called gatekeepers, giving administrators control over what an AI agent can modify or view, and when human approval is required. Agents operate with zero permissions by default. Cloudflare initially tested the platform internally, where employees across all departments used it daily. Over a recent 30-day period, internal staff built more than 4,000 applications and automations, while the sales department saved an estimated 10,000 hours on manual tasks like proposal creation.

Alongside the operating system, Cloudflare launched the Identity-Aware AI Gateway in beta, which integrates with identity providers like Okta or Entra to track both human and AI requests. A companion tool called AI Spend monitors usage patterns, while a User Insights tab flags anomalous activity by scoring sessions against a baseline. Specifically, it highlights any session costing more than twice the account's 95th percentile over the previous 30 days. This monitoring is crucial for preventing runaway costs; in one cited case, a rogue AI session left running by a customer's employee racked up a $30,000 bill before being detected.

For enterprise IT practitioners, these releases address the complex challenge of stitching together fragmented AI infrastructure. Instead of manually integrating disparate networking, security, and model routing components, administrators gain a unified platform to manage budgets, identities, and permissions. This infrastructure-first approach minimizes vendor lock-in and provides immediate visibility into token consumption, helping organizations avoid the budget overruns that have previously derailed enterprise AI deployments.