Cloudflare OS

• Preparing for a customer meeting by pulling account context from connected company systems and turning it into a slide deck to review with the team
• Building a live issue dashboard for a GitHub repository and sharing it with the team in real time
• Asking the agent to research a Gmail thread and related Drive documents, then draft a reply that is sent only after you approve it
• Fixing a bug in a connected GitHub repository and opening a pull request without a local development setup
• Turning a recurring report into a scheduled, mostly deterministic workflow that only calls a model where judgment is needed
• Giving non-engineering teams a governed place to use agents instead of handing out API keys to company systems

Cloudflare OS is less a product to sign up for than a starting point for building your company's own agent platform. Its strongest ideas are in governance: agents get only the resources a user hands them, every read is logged, sharing respects what the agent saw, and side effects queue up for later approval instead of forcing a choice between constant interruptions and auto-approve. That makes it worth a close look for security-minded organizations, especially ones already using Cloudflare Access and AI Gateway. The cost is ownership. You configure OAuth apps per integration, operate and upgrade an early-access codebase, and pay usage-based Workers and model costs. Teams that want a ready-made agent today are better served by Claude Cowork or Perplexity Computer until the managed version moves past its waitlist.

About Cloudflare OS

Cloudflare built Cloudflare OS for its own staff and open-sourced a rewritten second version on August 5, 2026. Each employee gets a browser-based workspace preloaded with skills and context the company curates, such as its terminology and procedures, so people do not re-explain them to a model for every task. The agent is a Code Mode agent: it does work by writing and running code in an isolated runtime, which lets it search and join data without pulling everything into the model's context. Outputs can be documents, slides, or spreadsheets that stay connected to live data, or full-stack apps (called Gadgets), each running in its own sandboxed Dynamic Worker with its own SQLite storage. Apps can be shared for real-time collaboration, or shared as Blueprints so colleagues get their own copy to modify with the agent. What sets it apart is the security model. Agents and apps start with access to nothing, and users grant specific resources, such as one GitHub repository or one Drive folder. Service-specific Gatekeepers hold the OAuth credentials, log what the agent reads, and handle any action with side effects. Instead of blocking on each approval, a Gatekeeper simulates the action so the agent can keep working, and the user approves or rejects the queued actions later, one by one or in bulk. Cloudflare OS also tracks what an agent has observed, so sharing its output cannot expose data the recipient could not see directly. Model calls go through Cloudflare AI Gateway, where admins choose models (the starter deployment defaults to Workers AI, with Anthropic, OpenAI, or Google optional), attribute spend, and set budgets. Updates announced on October 1, 2026 added GitHub repository work with commits and pull requests, Gmail drafting and sending, and exports to Excel, CSV, PDF, Markdown, and HTML. The tradeoffs: Cloudflare calls this an early-access release with rough edges, someone has to configure OAuth apps for each integration and keep the deployment updated, and costs land on your own Cloudflare and model provider bills. The managed option, where Cloudflare runs it for you, is only open as a waitlist.