Cloudflare announced the release of the new Cloudflare OS as an open-source project, a platform designed to enable members of an organization to build applications, automate business processes, and gain controlled access to internal systems through AI agents. Organizations can deploy the platform within their Cloudflare account, connect it to their own data, tools, and policies, and then customize it to reflect how they work.
The platform follows an initial internal release that has been used by thousands of Cloudflare employees since May 2026, including people outside the engineering teams. According to the company, the release was used to create documents and presentations, automate repetitive tasks, and build small applications for data visualization and everyday business support.
A Platform Combining Agents, Workspaces, and Applications
Cloudflare OS starts with a conversation inside the browser, but each workspace is linked to the context and skills the organization chooses to organize, such as terminology, procedures, and methods for carrying out repetitive work. A user can give the workspace a goal, allowing it to draw on the knowledge, tools, and data available within the organization to accomplish it.
The platform combines three main components: an agent workspace equipped with organizational context and skills and an isolated environment for writing and running code; a security and governance framework for accessing internal data and services; and a platform for building personal applications that can be customized and shared.
Workspaces can search and analyze information and create documents, presentations, and spreadsheets while keeping them connected to live data when necessary. They can also build connected applications that operate with their own interface, logic, and state, and support collaboration among multiple people. Work that follows a known sequence of steps can be converted into nearly deterministic workflows that run on demand, according to a schedule, or in response to an event in a connected system.
Access Starts with No Permissions
Cloudflare OS uses a model in which every agent and application starts with no permissions. The agent can request access to a specific resource, while the organization grants or denies it. Resources are made available to generated code through policy-constrained bindings, while credentials remain isolated from the agent and the code.
Code runs on the server inside a Dynamic Worker with general external connectivity disabled, while interface code runs inside an isolated frame in the browser. Neither can access the internet except through capabilities the organization explicitly grants.
The platform uses Gatekeepers, Workers-based services that sit between Cloudflare OS and external services. These services can restrict access to a specific repository on GitHub, allow issues to be read without source code, hide certain fields, enforce rate limits, or require approval before merging a pull request. The platform also supports existing Model Context Protocol servers through MCP Server Portals.
Cloudflare OS records the resources accessed by the agent and uses this record to verify the authorization of people attempting to open the workspace, interact with the agent, or view its outputs. Sensitive data that has been read can affect decisions about sharing work or sending requests to external services.
Shareable Applications and Controllable Models
Every file in Cloudflare OS can be a full application rather than a static document. The agent writes the interface portion that runs in the browser and the server portion that stores state and executes behavior. The server is loaded on demand inside a Dynamic Worker, while the application receives its own SQLite database through a Durable Object Facet.
The application itself can be shared for real-time collaboration, or its blueprint can be shared to create an independent copy. The new version includes the original code, but it does not include SQLite data, conversation history, credentials, or connected resources, giving each copy separate state and resources.
The platform supports any AI model and routes inference requests through Cloudflare AI Gateway. This layer allows organizations to define the available models, attribute spending to a person, team, or workspace, set budgets and rate limits, and determine what action to take when limits are reached.
Availability and Development Roadmap
Cloudflare OS is available today on GitHub through two repositories: a core repository and a repository for a reference deployment that reflects how Cloudflare operates it internally. It can be deployed in an organization’s Cloudflare account using its own Access policies, AI Gateway settings, data, and integrations. The interface can also be customized, and Gatekeepers and organization-specific features can be added without modifying the core product.
Cloudflare is working with its strategic partners Presidio and Happy Cog to customize the platform, connect it to internal systems, organize organizational skills and context, and configure security, model, and cost controls. The company said it is working to make Cloudflare OS available later within the Cloudflare dashboard as a fully managed product, add containers for development workflows, and provide workspaces within Slack and other conversational tools.