Cloudflare OS: Yet Another Agentic Work Environment?

Cloudflare OS: Yet Another Agentic Work Environment?

Promise kept, here it is: Cloudflare OS is now open source.

This platform brings together several internal projects. Built mainly on the Cloudflare Workers service (serverless functions), it is presented as a spiritual successor to Sandstorm.

From Sandstorm to Cloudflare OS, more than a family resemblance

Sandstorm first appeared in 2014. It was a server for hosting web applications. The project stood out for its level of granularity: every document had its own container. It ran into the technical limits of the infrastructures of the time (cold starts, memory consumption, etc.). The startup that backed it did not manage to monetize its model, which was based on providing enterprise features. In early 2017, it handed Sandstorm over to the community. Its team subsequently joined Cloudflare and began working on what would become Cloudflare Workers.

The functional units of Sandstorm were called “Grains.” With Cloudflare OS, they become “Gadgets”… and they drop containers in favor of a lighter isolation mechanism based on the V8 JavaScript engine.

Kenton Varda, cofounder of Sandstorm and today tech lead of Cloudflare Workers, demonstrated it at the AI Engineer World’s Fair at the end of June. The initiative was not yet officially named Cloudflare OS. But, according to him, it had become “seriously” internal enough not to be open sourced immediately — even though he had promised to do so when giving his talk.

The indie angle: empowering everyone to modify their apps

“Think of this environment as a kind of office suite,” Kenton Varda explained. But instead of documents, there are applications: the famed Gadgets. The sharing and access-control logic for these apps is implemented at the platform level. For each Gadget, the client code runs inside an iframe that has only a single communication channel: a POST to the parent frame. The link to the server-side code—which is instantiated as a durable object with its SQLite database—is accessed via an RPC session based on Cap’n Web, another open-source Cloudflare component.

Because the underlying runtime workerd is itself open, the entire stack can be run locally, Varda noted. He also explored the security model’s potential for vibe coding, or more precisely user-driven customization of apps by individuals running their own copies. He illustrated this with an example: by adding, through Claude, the ability to render arbitrary SVG diagrams and to center text within a slide-generation tool.

The B2B angle: codifying workflows in an agentive system

Kenton did not dwell at length on another aspect: linking these apps to external services. Cloudflare now handles that, since the platform is open. In a sense, this is a v2. The first version was launched internally in May. It allowed running workflows codified as skills via a web UI. These had been developed in advance, based on a mapping of tasks that employees wished to automate. The results appeared in a side panel within the browser, with no interaction with the rest of the local environment.

Towards a “factory of connectors” for Cloudflare OS

In Cloudflare OS terminology, these ad hoc MCP servers are “Gatekeepers.” They house the logic of each service and ensure the enforcement of security policies when resources are shared. They can be the applications themselves (with databases, chat histories, authenticators, etc.) or blueprints (code only).

Gatekeepers are capable of “simulating” the outcome of actions. They can be validated in batches at the end of the chain, without the agent interrupting its work at every step.

Cloudflare is working on a form of a “software factory” that will produce Gatekeepers from API documentation. There are currently Gatekeepers for Confluence, GitHub, Google (Gmail, Docs, Sheets, Calendar, BigQuery), Home Assistant, Linear, Notion, Slack, Spotify, Supabase and ZoomInfo. Cloudflare adds a ready-to-use implementation model. It enables exploiting a pinned version without modifying the source code, while still having control over identities, routing, integrations, inference, observability and branding.

Dawn Liphardt

Dawn Liphardt

I'm Dawn Liphardt, the founder and lead writer of this publication. With a background in philosophy and a deep interest in the social impact of technology, I started this platform to explore how innovation shapes — and sometimes disrupts — the world we live in. My work focuses on critical, human-centered storytelling at the frontier of artificial intelligence and emerging tech.

© 2025 Dawn Liphardt