Cloudflare's internal AI productivity platform, open-sourced. Agents build "gadgets" — private per-user apps written from a prompt — and reach company systems through Gatekeepers, a capability-based permission layer.
Cloudflare runs this in-house across engineering, sales, and everything between. What's on GitHub is v2, a full rewrite, Apache-2.0, explicitly meant to be forked into Your Company OS rather than used as-is.
Think of an office suite where every file is its own app. Ask the agent for a slide deck or a collaborative whiteboard and it writes one — a "gadget" — that only you run, in its own sandbox, and that you can keep changing by prompt instead of filing a feature request. Share the live instance for real-time multiplayer, or share the code as a Blueprint so everyone gets their own copy.
The Gatekeeper layer is the part worth stealing. Each external service gets a Worker that owns OAuth, narrows access to the single resource you introduced, and logs every call. For side-effecting actions it doesn't block — it simulates the result so the agent keeps working, then you approve the queue in bulk later. That's an honest answer to --dangerously-skip-permissions.
Built by the Workers team, on Durable Objects and Dynamic Worker Facets — runtime features added specifically for this, so the source doubles as a reference for how Workers are meant to be used. The agent itself is a Code Mode agent. pnpm run-local runs the whole stack on workerd; one-click deploy at os.cloudflare.app.