Cloudflare OS: An AI productivity environment

Cloudflare OS is an open-source "operating system" for AI productivity built by the Cloudflare Workers team. It provides an agent chat UI, sandboxed application development (where agents build "gadgets" — small personal apps), and a capability-based security framework called Gatekeepers. Originally developed for internal use, it's now open-sourced so other companies can run their own instance.

The architecture is built entirely on Cloudflare Workers and Durable Objects. Every workspace is its own Durable Object, every Gadget runs in a Dynamic Worker Facet, and Gatekeepers install facets into workspaces to manage access to remote services. The frontend and backend communicate via Cap'n Web RPC over WebSockets, with Yjs (V2 encoding) syncing code changes between clients and agents in real time.

Gatekeepers are the security innovation: they wrap external services behind a capability-based API, logging every action and providing simulated outcomes for async human-in-the-loop approval. Instead of making the agent stop and wait for approval on every side-effecting action, Gatekeepers simulate results locally so the agent can proceed, queuing actions for bulk human approval later. This solves the "agent stuck on first approval" problem that drives users to dangerous auto-approve modes.

The agent uses a Code Mode approach — it performs tasks by writing and immediately executing JavaScript snippets via Dynamic Workers, rather than calling a fixed set of predefined tools. This makes every Gadget's Cap'n Web API automatically available to the agent without additional MCP integration.

Gadgets are per-user, sandboxed application instances — like Google Docs but where each "doc" can be a completely custom app built by AI. Blueprints are shareable gadget templates (.gadget binary format). The permission system uses capability-based introductions: agents and gadgets start with access to nothing and must be explicitly introduced to each resource they need.