Cloudflare has released Cloudflare OS, an open-source platform for building and running AI agents and full-stack apps, and it is available today on GitHub. The pitch is direct: instead of wiring together a model provider, a runtime, a database, and an auth layer yourself, you deploy one governed operating layer into your own Cloudflare account and start shipping agents and apps in minutes. For anyone building tools, automations, or products on top of AI, this is a serious new option in a crowded field, and its open-source, self-hostable license is what sets it apart.
What Cloudflare OS Actually Ships
Cloudflare OS bundles three things that builders normally assemble by hand: agent workspaces, a full-stack app runtime, and a security and governance layer. The core platform is on GitHub, alongside a starter repository you can deploy into a Cloudflare account without standing up any infrastructure of your own.
The agent workspaces handle research and analysis grounded in your own systems and terminology, generate documents, slides, and spreadsheets, and run deterministic workflow automation. The app side lets you build full-stack applications where each app runs as an isolated unit with its own database. The governance layer is the piece most agent stacks skip: every agent starts with zero access and must request specific resources, which are then mediated by policy.
How It Is Built: Workers, Gatekeepers, and Dynamic Workers
Cloudflare OS is assembled from Cloudflare's existing primitives rather than a new bespoke stack. App code runs on Cloudflare Workers, with each app spun up as a Dynamic Worker, a lightweight V8 isolate. Every app gets its own isolated SQLite database through Durable Object facets, so one app cannot read another's data by default. Client and server talk over Cap'n Web, Cloudflare's object-capability RPC system, and client code runs inside a browser sandbox.
The security model centers on what Cloudflare calls Gatekeepers. A Gatekeeper is a service-specific Worker that sits between the platform and an external system: it holds the credentials, enforces resource-level policies, applies rate limits, records what data an agent touched, and can mask sensitive fields. The canonical example is granting an agent access to a single GitHub repository while blocking certain file types, rather than handing it a broad token. Model calls are routed through Cloudflare AI Gateway, which keeps the platform model-agnostic and adds cost controls on top.
Cloudflare OS vs Other Ways to Build Agents
The agent-platform space has split into hosted, closed offerings and self-hosted, open ones. Cloudflare OS lands firmly in the open, self-hostable camp, but with managed infrastructure underneath. Here is how it compares to the patterns builders are using now.
| Approach | Hosting | Model choice | Governance built in | License |
|---|---|---|---|---|
| Cloudflare OS | Your Cloudflare account | Any, via AI Gateway | Yes (Gatekeepers, zero-access default) | Open source |
| Closed hosted agent platforms | Vendor cloud | Usually vendor-locked | Varies, vendor-defined | Proprietary |
| DIY frameworks (glue code) | Self-managed servers | Any | No, you build it | Mixed |
| Managed agent runtimes | Vendor cloud | Often single provider | Partial | Proprietary |
Cloudflare has been building toward this for months. In May 2026 it shipped Cloudflare Environments for Claude Managed Agents, a way to run Anthropic's agents on Cloudflare's network while orchestration stayed inside Claude. Cloudflare OS generalizes that idea into an open platform that is not tied to one lab's agent framework, which is the substantive shift. It is a close cousin to open agent workspaces like Block's Buzz, but it leans harder on Cloudflare's edge runtime and its built-in permission model.
Why It Matters for Builders and Creators
If you build websites, apps, internal tools, or content pipelines, the hardest part of adding AI agents is rarely the model call. It is the plumbing around it: where the agent runs, what it is allowed to touch, how you stop it from leaking data, and how you keep model costs from spiraling. Cloudflare OS ships answers to all four as defaults rather than homework.
The zero-access-by-default posture matters for anyone letting an agent act on real systems. Recent incidents, from an AI agent attacking Hugging Face to memory-exfiltration bugs, have made clear that broad tokens plus autonomous agents is a bad combination. A platform that forces per-resource requests and records every data access is a meaningfully safer starting point than glue code that grants the agent everything. For solo builders and small teams, the fact that it deploys onto Cloudflare's managed runtime means you get that governance without running your own servers.
How to Deploy Cloudflare OS
Getting a working instance follows a short path, and Cloudflare's own framing is that it takes a few minutes from the starter repo.
1. Clone the starter. Start from the cloudflare-os-starter repository rather than the core repo. It is the deployment template wired for a fresh Cloudflare account.
2. Deploy into your account. Push the starter into your own Cloudflare account. Apps come up as Dynamic Workers, each with its own SQLite database, so you are not provisioning infrastructure separately.
3. Connect a model through AI Gateway. Because the platform is model-agnostic, you point it at whatever provider you want through AI Gateway and set cost controls there.
4. Add Gatekeepers for anything external. For each outside system an agent needs, GitHub, an internal API, a data store, add a Gatekeeper Worker that scopes exactly what the agent can do and logs what it accesses.
5. Build or share apps. Build full-stack apps in the workspace, then share them directly or publish them as blueprints others can customize. Cloudflare named Presidio and Happy Cog as launch partners for teams that want help integrating internal systems.
What to Watch
Two things will decide how far this goes. First, community traction: the launch landed near the top of Hacker News, and an open license means forks, third-party Gatekeepers, and blueprint libraries could form quickly. Second, the model-agnostic promise: AI Gateway routing is the lever that keeps builders from being locked to one lab, and how cleanly that holds up in practice, across providers and costs, is the real test. For builders comparing options, Cloudflare OS is worth a weekend deploy before committing a production agent stack to any single vendor.
Frequently asked questions
Is Cloudflare OS free and open source?
The platform is open source and available on GitHub today, and you deploy it into your own Cloudflare account. Cloudflare did not publish separate pricing for the platform itself in the announcement; the usual Cloudflare Workers, Durable Objects, and AI Gateway usage costs apply to what you run on it.
What can I actually build with it?
Two things: AI agents that do research, generate documents, and run workflow automation grounded in your own systems, and full-stack applications that run as isolated Dynamic Workers with their own databases. Apps can be shared directly or published as blueprints others customize.
Do I have to use a specific AI model?
No. Cloudflare OS routes model calls through Cloudflare AI Gateway and is model-agnostic, so you can use any supported provider and manage cost controls centrally rather than being locked to one vendor.
How is it different from Cloudflare's Claude Managed Agents?
The May 2026 Claude Managed Agents work ran Anthropic's agents on Cloudflare while orchestration stayed inside Claude. Cloudflare OS is a general open platform that is not tied to one lab's agent framework, and it adds the app runtime and the Gatekeeper governance model.
What are Gatekeepers?
A Gatekeeper is a service-specific Worker that mediates an agent's access to an external system. It holds credentials, enforces per-resource policies, applies rate limits, logs data access, and can mask sensitive fields, so an agent gets scoped access (one repo, certain file types) instead of a broad token.
Is this only for large enterprises?
No. While the governance features suit organizations, the platform deploys from a starter repo into any Cloudflare account in minutes, which puts it within reach of solo builders and small teams who want managed infrastructure and built-in permissions without running their own servers.
