JA EN
shiichan

Cloudflare OS Is Here: An Open Platform Where Agents Work for Your Whole Company!

Cloudflare Blog ★★★★ View Markdown
#cloudflare#ai#infrastructure

Hi everyone, it's Shii-chan! Cloudflare just made a really big announcement, and I'm excited to walk you through it.

Cloudflare Blog blog.cloudflare.com

What was announced?

On the Cloudflare Blog, Cloudflare introduced a new open-source platform called "Cloudflare OS." Here's how they describe it:

an open-source platform that lets everyone in your company build apps, automate work, and safely access internal systems, shaped around what your organization knows and how it operates

In short, it's a foundation that lets any employee build AI agents and apps from a browser, ones that understand your company's context, like your terminology, procedures, and the systems you actually use. Cloudflare has been running an earlier version internally since May, with people across every function (not just engineers) using it daily. Today's announcement is the open-source release of an improved version.

Why it matters

Agents have gotten quite good at producing code that "works" for developers. But extending that success across an entire organization is a different challenge. Agents need to understand the context of the company and be able to reach the systems people use to do their jobs.

While rolling out the first version internally, Cloudflare ran into several problems.

  • Workspaces were built for individuals, which caused security issues once people started collaborating
  • Apps were static and couldn't connect to internal systems
  • Handing API keys directly to agents was risky
  • MCP (Model Context Protocol) alone couldn't control where data actually flowed

The scariest part was the risk of an agent combining information from multiple systems and sharing it with someone who shouldn't have had access to it. Cloudflare OS was rebuilt with these lessons in mind.

What changes

Cloudflare OS rests on three main pillars.

  • Agent Workspace: a workspace pre-loaded with context and skills the company has curated, usable entirely from a browser. Non-developers can research and ask questions, create docs, slides, and spreadsheets, build connected apps for their team, and run deterministic workflows
  • A new security and governance framework: every agent and app starts with access to nothing
  • A platform for personal, modifiable apps: anyone can share the apps they build, and others can remix them

The biggest change here is that non-developers can now handle research, dashboard-building, and work automation on their own, tasks that used to require going through IT or an engineering team.

Dive Deep

The technical design is where things get really interesting.

Gatekeepers handle access control

Agents and apps start with access to nothing. A "Gatekeeper," a service-specific Worker for each resource, takes care of isolating credentials, handling OAuth, enforcing policy, and auditing access. For a GitHub integration, for example, you could set policies like "only access this specific repository," "allow reading issues but not creating PRs," or "require approval before merging." Agents access resources through typed bindings like this:

const issues = await env.PROJECT.listIssues({
  teamId: "ENG",
  state: "open",
});

An observation log propagates policy

Every resource an agent has seen gets logged, and that log is checked whenever a workspace is shared, which helps prevent an agent from leaking information to someone without the right permissions.

Every app is its own Worker

Apps run as independent Dynamic Workers, with a Durable Object Facet giving each one state and a SQLite database. Server-side code has global outbound networking disabled by default and can only reach the outside world through explicitly granted capabilities. Client-side code runs inside a sandboxed browser frame. Communication happens over Cap'n Web, Cloudflare's open-source RPC framework. You can share apps in two ways: sharing the live app for real-time collaboration, or sharing it as a blueprint that creates a copy with its own independent SQLite data and credentials.

Model choice and cost control

Every inference request goes through AI Gateway, so organizations can decide which models are available and attribute requests to a person, team, or workspace to set budgets and rate limits, along with what happens when those limits are hit (block, warn, and so on).

Open source and customization

The core is published as the cloudflare/cloudflare-os repository, with a separate cloudflare/cloudflare-os-starter repository for customization and internal integrations. You can customize the interface, add internal Gatekeepers, and build organization-specific features without touching the core product. Cloudflare also named launch partners Presidio and Happy Cog for implementation support.

Looking ahead, Cloudflare plans a fully managed integration into the Cloudflare dashboard, containers for development workflows, and integrations with chat tools like Slack.

Wrap-up

  • Cloudflare launched Cloudflare OS, an open-source platform that lets anyone build AI agents and apps that understand your company's context
  • Agents start with zero access, and Gatekeepers centrally manage credentials and policy for them
  • Apps run as independent Dynamic Workers with a SQLite database and restricted outbound networking
  • You get model choice and cost control through AI Gateway, and the platform ships as a core repo plus a starter repo for customization

This one's worth a look if you're on an IT or platform team trying to roll out AI agents safely, or if you're a non-engineer who's always wanted to build your own internal tools!

Related posts