What Is Cloudflare OS? Architecture and Components Explained

Published: Last updated:

In brief: Cloudflare OS is an open-source enterprise AI agent platform released by Cloudflare in August 2026 under the Apache 2.0 licence. It is deployed to your own Cloudflare account, where it connects institutional knowledge and business systems so employees can use AI agents and workplace apps in a browser.

Its defining feature is that security is built into the platform rather than left to each user. Every agent starts with zero permissions, while a Gatekeeper brokers and records each operation on an internal system.

Cloudflare OS at a glance

The name includes “OS”, but this is not an operating system like Windows or Linux. It means a workplace platform for running AI agents inside an organisation.

Basic information about Cloudflare OS
NameCloudflare OS, pronounced “Cloudflare O-S”
ProviderCloudflare, Inc., a US company known for CDN and security services
ReleasedReleased as open source on 4 August 2026, US time
LicenceApache 2.0; the software licence fee is $0
DeliverySelf-deployed to your Cloudflare account; a managed Dashboard edition is planned, with pricing not yet announced
AccessEmployees open the workspace in a browser
Source codecloudflare/cloudflare-os; official starter: cloudflare-os-starter

Cloudflare says several thousand of its own employees already use the platform; that figure is Cloudflare’s own report. It is also a newly released Early Access product. We return to this under adoption considerations.

Why it exists: three limits of general-purpose AI

Giving a company a general-purpose AI tool such as ChatGPT does not automatically make day-to-day work faster. Organisations commonly encounter three barriers.

1. AI does not know how your company works

A general-purpose model does not know internal terminology, procedures or system conventions. If employees must explain these from scratch every time, the time saved in real work remains limited.

2. It cannot reach internal systems

AI needs access to systems such as a CRM or database to create meaningful operational value. Giving API keys directly to employees or agents is risky, however, and makes it difficult to audit who did what.

3. Permission boundaries become unclear

Managing which tools an AI may use is not enough. Unless the organisation can trace which data it read, sharing an AI-produced artefact internally can itself become a route for unintended disclosure.

Cloudflare OS is designed to embed controls in the platform instead of relying on each user’s security awareness to resolve these problems.

How it works: five core components

Cloudflare OS is both an internal AI workspace and a platform for creating AI applications, deployed within the organisation’s own Cloudflare account. It has five principal components. A flat list hides how they connect, so the flow is shown first as three layers.

The five components as three layers

Read from the employee-facing surface, through the control points, to the destinations. That order shows who mediates each request.

What employees use

Workspace The browser place to work with AI. Employees enter after Access authentication
Blueprint / Gadget Reusable designs and building blocks for apps and workflows

Actions and model calls must pass a control point

Control points

Gatekeeper Allows, denies and records operations on internal systems
AI Gateway The window for model calls and cost control
MCP Server Portal The catalogue of internal integrations

Only approved destinations are reached

Destinations

Internal systems Reached through Gatekeeper and MCP
AI models Called through AI Gateway
Cloudflare OS components and their roles
ComponentRole
Workspace RoleA browser environment where employees work with AI agents. Agents generate and run code in isolated execution environments, while sessions and output files are retained.
Gatekeeper RoleThe control point between AI and internal systems. A service-specific Worker allows, denies or holds each operation for human approval and records what was read.
AI Gateway RoleThe route for every AI inference request. It centralises model management, cost visibility, budgets and rate limits.
MCP Server Portal RoleAn organisational catalogue for publishing and managing internal MCP servers and the integrations available to agents.
Blueprint / Gadget RoleReusable designs and building blocks for applications and workflows. Sharing them turns an effective way of working into an organisational asset.

Typical request flow

An employee first authenticates through Cloudflare Access and opens the workspace. When an AI agent needs to act on an internal system, it must go through Gatekeeper. Every model request also passes through AI Gateway, so the organisation can see and control which models are used and in what volume.

The platform is not guaranteed to be no-code; custom integrations require development. See the official repository for details.

Security model: begin with zero permissions

Cloudflare OS starts from a deny-by-default position and opens only the access a task needs, rather than allowing everything first and attempting to narrow it later.

  • Nothing is visible at first — Every agent and application starts with zero access. Resources are requested, and an administrator decides whether to allow or deny them.
  • Gatekeeper brokers operations — Controls such as read-only access, masking particular fields and requiring human approval before execution can be applied to individual operations.
  • Policy follows the information read — Referenced resources are recorded and linked to outputs. The platform can verify whether a viewer may access the source data and restrict actions such as external transmission after confidential data has been read.

These controls do not eliminate risk. Prompt injection and permission-design errors remain possible, so a limited, read-oriented rollout is the prudent starting point.

What teams can do with it

Conversations in the workspace can produce documents, applications and workflows. Representative uses include the following; these are illustrations, not guaranteed outcomes.

  • Sales — Research before meetings, draft proposals and create CRM summary reports
  • Customer support — Search enquiries and knowledge bases together, draft responses and FAQs, and analyse trends
  • Leadership and planning — Generate cross-functional dashboards and management-meeting summaries
  • Human resources — Answer questions about internal policies, coordinate recruitment schedules and produce onboarding materials
  • IT — Design Gatekeeper and MCP integrations, operate access policies and monitor logs
  • Corporate operations — Check expense summaries, extract contract highlights and turn internal requests into workflows

Adoption considerations

  • It is an Early Access product — Cloudflare OS was released only in August 2026, and the official starter asks adopters to pin releases and review changes. A staged rollout beginning with a limited pilot is more realistic than an organisation-wide launch.
  • Deployment alone does not create value — Installing the open-source software will not change work by itself. Teams must select suitable workflows, implement Gatekeeper and MCP connections, and train employees so the new approach takes hold.
  • A $0 licence does not mean $0 to operate — Workers, Access, AI inference and related services carry separate charges. See the pricing and costs guide.

We can help you assess whether Cloudflare OS fits your organisation, beginning with workflow selection.

Discuss adoption