Cloudflare OS

// Overview

Cloudflare OS is an "operating system" for AI productivity that Cloudflare originally built for its own staff, from engineering through sales, and then released as open source. The name is used in two senses: an operating system for a company to be productive with AI safely enough that the security team can sleep, and an operating system for AI workloads in the sense that a traditional OS manages compute. It provides three things — an agent chat UI preloaded with knowledge of how your company operates, sandboxed application development, and a security framework called Gatekeepers. The stated intent is not that you run Cloudflare's instance but that you fork it into "Your Company OS".

The unit of software is a gadget: a private instance of a productivity app that belongs to one user. When you create a slide deck, the system does not call out to shared SaaS — it creates your own copy of the slide-deck software in its own sandbox. Two things follow. The app cannot leak your slides through a bug in someone else's instance, because the sandbox mediates all access; and you can freely modify the code, asking an agent to add a missing feature, because the blast radius stays inside your own instance.

Gatekeepers are the security layer, described as supercharged MCP servers. Introducing an agent or gadget to an external resource creates a Gatekeeper for that service: it wraps the native API in a clean Cap'n Web API, handles authorization such as OAuth, narrows access to the specific resource the user intended, logs every action for review, and puts a human in the loop on side effects. Its notable trick is asynchronous approval — the Gatekeeper simulates the outcome locally so the agent can keep queueing work, returning simulated results if the agent reads back, and the user approves or rejects in bulk later instead of blocking the agent mid-task. Each Gatekeeper is a separate Worker. The repository is version 2, a complete rewrite, and Cloudflare labels the August 2026 release early access with rough edges.

// What it does

• An agent chat UI grounded in company context, with GitHub and Google integrations
• Gadgets: per-user private instances of productivity apps, each in its own sandbox and freely modifiable by prompting an agent
• Gatekeepers, a capability-based security layer that wraps each external service in a Cap'n Web API, handles OAuth and narrows access to the intended resource
• Asynchronous human-in-the-loop approval — actions are simulated so the agent keeps working, and the user approves or rejects in bulk later
• Full action logging per gadget and agent for after-the-fact review
• Runs entirely on Cloudflare Workers, deployable into your own Cloudflare account or runnable locally on wrangler and workerd

// Getting started

The whole stack runs locally on wrangler and workerd, which is the fastest way to see what it does; deploying to your own Cloudflare account is the other path.

pnpm run-local
# then visit http://localhost:8787

https://os.cloudflare.app/deploy

Make slides for my upcoming meeting with a customer.
Make a collaborative whiteboard app.
Make a tic tac toe game.

Make an issue dashboard for this GitHub repo.
Fix the typos in this Google Doc.

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Give non-engineers a safe way to build and modify their own internal tools by prompting an agent
• Ground an internal assistant in company systems while keeping every external action logged and approvable
• Replace shared SaaS instances of simple productivity apps with per-user sandboxed copies
• Fork a working reference implementation of capability-based agent security rather than designing one from scratch

// How Cloudflare OS compares

Cloudflare OS alongside other open-source agent frameworks & builders tools AI/TLDR tracks, ranked by GitHub stars.