An open-source platform records every resource accessed by AI agents, enforces permission-based access, and enables secure collaboration while maintaining an auditable history of agent activity.
Cloudflare has open-sourced Cloudflare OS, an AI agent platform designed to improve security, governance, and collaboration in enterprise AI workflows. The platform maintains a persistent record of every resource an AI agent reads and associates that record with the agent’s outputs. Before another user can access those outputs, Cloudflare OS verifies that they have permission to view the underlying data, preventing unauthorised information sharing.
While existing tool integration protocols only describe how an agent accesses external resources, Cloudflare OS also logs which data has been accessed by the agent. The observation log follows the agent throughout the workflow and allows organisations to enforce their policies based on the history of data the agent has accessed. An agent that has accessed sensitive data will be prohibited from sending results outside, delegated tasks to other agents, or communication with forbidden destinations.
The platform takes a zero trust approach to handling credentials. AI agents never have direct access to credentials or authentication tokens. The service, called Gatekeeper, handles authentication, credential storage, access policies enforcement, and logging all permitted operations. The agent only communicates with managed interfaces that allow administrators to establish detailed permissions like repository access, data masking, approvals, and rate limits.
The platform enables workspaces to build applications whose client and server components run in separate environments. Developers can use the Cloudflare remote procedure call framework to create tools that are accessible not only by users but also by AI agents, with the application data, credentials, and environments remaining isolated from one another.
The project is currently available as an open-source project on GitHub. Managed deployments, development containers, and collaboration workspaces are planned for future releases.
