Cloudflare OS: Here’s What’s Inside the Open-Source AI Agent Platform
Cloudflare first built Cloudflare OS for internal use, then opened a rebuilt version for other organizations to deploy and connect to their own systems. The goal is to make agentic work usable beyond engineering while baking security into the platform itself. The system has three main parts: an agent workspace grounded in company context with an isolated runtime for code execution; a security and governance layer between agents and internal systems; and a layer for modifiable apps that can turn chats into documents, workflows, or small full-stack apps. A key design choice is avoiding broad API keys. Agents start with no access, request specific resources, and a service-specific Gatekeeper mediates access without exposing credentials to the agent. The platform also logs observations and checks user access before showing workspace contents or agent output. Each app runs as a Cloudflare Worker using Cloudflare-built infrastructure and RPC. The open-source code is available, but the runtime remains Cloudflare-controlled.
