Early-access release on GitHub packages the internal platform used across thousands of employees, as Cloudflare pushes governed agent workflows, app building and edge-based deployment for enterprises.
Cloudflare has released a rebuilt version of its internal AI platform, Cloudflare OS, as open-source software under the Apache License 2.0, positioning it as a way for enterprises to deploy autonomous agents and custom applications without broadly exposing sensitive internal systems. The company says the same foundation is used internally across its thousands of employees. Cloudflare OS is not a traditional operating system. It is designed as a control layer for productive and safe enterprise AI use, managing AI workloads in a way Cloudflare likens to how an operating system manages processes. The platform targets developers building autonomous apps and workflows on Cloudflare's edge infrastructure, underscoring the company's push into agent infrastructure. The rebuilt platform combines an agent workspace, a new security and governance framework, and a layer for personal, modifiable apps. The workspace gives agents curated company context and skills inside an isolated runtime where they can write and run code. The app layer lets users turn chat prompts into documents, workflows or small full-stack applications. Cloudflare said the redesign addresses weaknesses in the earlier internal version, where even deterministic tasks consumed AI model tokens and where sharing outputs risked exposing data to people without the right permissions. In the new system, deterministic tasks run as conventional code and AI models are used only when judgment is needed. Users can describe goals in natural language and generate documents, spreadsheets, dashboards or full-stack business apps, while workflows can run on demand, on schedules or in response to internal events. Security is centered on Gatekeeper, a service-specific Cloudflare Worker (serverless application on Cloudflare) that mediates access between Cloudflare OS and external services. Agents begin with no access, and users or administrators grant only specific permissions to data and operations. Cloudflare's CIO Sam Rhea said the design avoids handing broad, long-lived API keys to people or agents. Instead, credentials stay away from the agent, Gatekeeper brokers each request, and the system logs every resource an AI observes while checking a person's rights before they can open a workspace or view generated output. That model is meant to extend governance beyond the moment data is sent to an AI. A dashboard built from confidential information, for example, will not display unless the viewer has access to the source data. Applications built in Cloudflare OS are Cloudflare Workers running on Dynamic Workers and Durable Object Facets, and they communicate with clients over Cap'n Web, Cloudflare's open-source object-capability RPC system. Shared apps can also be cloned as Blueprints, which include program logic but exclude saved data, conversation history and authentication credentials. The launch comes with Cloudflare Wallets and cloudflare.pay, aimed at identity and payment problems as AI agents begin transacting online. Each Cloudflare account will receive a unique web address as a persistent identity that can be extended to specific agents. The wallet system is designed to store, receive and manage stablecoins, and account holders can create Virtual Wallets for agents with limits such as daily caps, per-transaction caps and approved merchant lists. CEO Matthew Prince said in May that every one of Cloudflare's thousands of employees had access to an earlier version, with staff outside engineering using it for drafting documents, building slides and automating repeatable work. He said, "The Internet is shifting from human-driven browsing to agent-driven commerce, and the infrastructure needs to keep up," and added, "When an agent shows up at your door, you need to know who sent it." Cloudflare OS is available in early access on GitHub, and the company plans a fully managed service through its administration console as well as integrations including Slack. The release's open-source structure lets organizations deploy and customize the platform on their own infrastructure, although its agents, apps and governance framework run on Cloudflare's edge when customers use the managed service. The hybrid model may invite scrutiny because the Apache 2.0 license opens the code while the runtime environment remains under centralized control. Cloudflare argues the Gatekeeper approach improves on Model Context Protocol by logging what an agent has actually seen and checking access before any output is displayed, embedding access control into the platform layer rather than leaving it to each app builder.