Cloudflare OS Review (2026): Pricing, Features & Honest Verdict

Cloudflare OS is a free, Apache-2.0 AI workspace you deploy into your own Cloudflare account, giving every employee agents that reach internal systems through policy-enforcing proxies. The software costs nothing. The platform underneath it does, and Cloudflare never adds that bill up for you. Best for: teams already running on Cloudflare. Price: Free, self-hosted (free plan: yes). Rating: 7.5/10

What is Cloudflare OS?

Cloudflare OS is Cloudflare’s open source answer to a question most companies are currently failing: how do you give the whole organisation AI agents that can touch real internal systems without handing every employee a set of production credentials? Coding agents solved this for engineers. Cloudflare’s framing is blunt about the rest: “Bringing the same leverage to the rest of the organization is a harder problem.”

It launched on 4 August 2026 and took the number one spot on Product Hunt on 6 August. It is not a hosted product you sign up for. You clone a repository, deploy it into your own Cloudflare account, and wire it to your own Access policies, AI Gateway configuration, data and integrations. Cloudflare publishes two repositories: the core, and an example deployment mirroring how Cloudflare runs it internally. The deployment repo consumes the core without patching it, which is the right separation and means upgrades will not fight your customisations.

What Are Cloudflare OS’s Key Features?

Gatekeepers hold the credentials, not the agents

This is the load-bearing idea. A Gatekeeper is a service-specific Worker that sits between an agent and an external system. It holds the credentials separately from the agent and enforces fine-grained policy: repository-level GitHub access, field masking, approval workflows before a write lands. The agent never sees the token. Most internal AI rollouts fail on exactly this problem, and most competing products solve it by asking you to trust the model. Cloudflare solved it with a proxy, which is a much older and much better answer.

Personal apps built on Durable Objects

Agents can build full-stack applications for individual users using Dynamic Workers and Durable Object Facets with SQLite databases behind them. You can share a finished app directly, or share a blueprint so someone else can fork and modify it. The pitch is that a colleague who wants a change edits the app with AI instead of filing a ticket against you. Whether that survives contact with a real company is untested, but the architecture is sound and the storage layer is a real database rather than a JSON blob.

MCP Server Portals and a shared skills library

If your organisation already runs MCP servers, Cloudflare OS connects to them through MCP Server Portals rather than making you rebuild. Alongside that sits a shared context and skills library holding your terminology, procedures and best practices, so every agent starts grounded in how your company actually talks about its own work. This is the part that partners get paid to build out, and it is also the part that determines whether the deployment is useful or a toy.

Every inference call runs through AI Gateway

Cloudflare OS works with any model, and every call routes through Cloudflare AI Gateway. That gives you one place to decide which models are available, which model handles which job, and how spend is attributed per person and per team, plus budgets and rate limits. For an organisation worried about an uncapped AI bill, per-person cost attribution built into the platform is a stronger control than a spreadsheet and a policy memo.

How Much Does Cloudflare OS Cost?

The software is free. It is Apache-2.0 licensed and available on GitHub, and there is no licence fee, no seat cost and no paid tier of Cloudflare OS itself. What you pay for is the Cloudflare platform it runs on, and that is where a real number lives.

Four products carry the bill. Workers Paid starts at $5 a month minimum for an account, including 10 million requests and 30 million CPU-milliseconds, then $0.30 per additional million requests and $0.02 per additional million CPU-milliseconds, with no charge for bandwidth. Durable Objects on the paid plan include 1 million requests and 400,000 GB-seconds a month, then $0.15 per million requests, $12.50 per million GB-seconds, $1.00 per million SQLite rows written after the first 50 million, and $0.20 per GB-month of storage after the first 5 GB. AI Gateway’s core features are free on every plan. And Cloudflare Access, which controls who gets in the door at all, is where the number jumps.

Cloudflare positions its Zero Trust Free plan for “teams under 50 users” at $0 forever, and its Pay-as-you-go plan at $7 per user per month for “teams over 50 users.” So a 40-person company can run Cloudflare OS for about $5 a month plus model spend. A 200-person company pays $1,400 a month for Access alone before a single token is generated. Crossing 50 people is the most expensive thing that will happen to this deployment, and nothing on the launch page mentions it.

What Is the Real Total Cost of Ownership?

Model inference is the other half and it is not Cloudflare’s to price. You bring your own provider, and you pay that provider. The one Cloudflare-specific charge is Unified Billing, which adds a 5% fee on credits purchased through it: buy $100 of credit and you are charged $105. Route through AI Gateway with your own provider keys and you avoid that entirely, which is worth knowing before you set it up the convenient way.

Two smaller line items catch people out. Persistent logs are capped at 100,000 total across all gateways on Workers Free versus 10 million per gateway on Workers Paid, so an organisation-wide deployment on the free plan will start dropping its audit trail almost immediately. And Guardrails, the safety filter, evaluates prompts and responses using a Llama Guard model on Workers AI, billed as token inference. Turning safety on costs money that scales with how much your company talks to its agents.

Then there is the cost nobody invoices you for. This is self-hosted software with no tagged releases in its repository, so you track the main branch. Cloudflare’s own answer to that is a partner network: Presidio and Happy Cog will customise and roll it out for you. That is a consulting engagement with a consulting price, and for most companies below a few hundred people it will dwarf every platform figure above.

Is the Project Actually Healthy?

Yes, and the numbers are checkable in a minute. The core repository was created on 15 April 2026, has 9,476 stars and 1,110 forks, was last pushed on the day we checked, and is written in TypeScript under Apache-2.0. It has 19 contributors. The top one is kentonv with 382 commits, which is Kenton Varda, the person who built Cloudflare Workers in the first place. The second is one of the engineers who wrote the launch post. This is not a marketing repo with a README and a logo.

The caveats are also visible. There are 111 open issues on a project four and a half months old. There are zero tagged releases, so there is no version to pin and no changelog to read before you upgrade a system your whole company logs into every morning. The starter repository, at 274 stars and 98 forks, has not been touched since 19 August. Cloudflare says a fully managed dashboard version is coming, along with containers and Slack workspaces. Deploying today means running an unversioned main branch until that lands.

The strongest evidence in its favour is that Cloudflare runs it. The company reports that “Thousands of people across every function, many of them outside of engineering, use it every day.” Dogfooding at that scale is a better signal than any customer logo wall.

Who is Cloudflare OS Best For?

Use Cloudflare OS if you already run on Cloudflare, you have engineers who are comfortable with Workers and Wrangler, and you want to give a non-engineering department real AI leverage without buying a per-seat product for every person in it. Under 50 people, the platform cost is close to nothing and the value is immediate.

Use it if governance is the blocker rather than capability. Gatekeepers, Access and per-person cost attribution in AI Gateway are a stronger compliance story than almost anything you can buy, and you can read the code that enforces it.

Skip Cloudflare OS if you do not have platform engineering capacity. This is infrastructure you operate, not software you subscribe to, and running it on an untagged main branch needs someone whose job that is. Skip it if you are not on Cloudflare, because the entire value proposition assumes Workers, Durable Objects, Access and AI Gateway. And wait for the managed version if you want a product rather than a project.

Best Cloudflare OS Alternatives

Glean is the enterprise version of this idea, sold the enterprise way. It does workplace search and assistants across your internal tools, with connectors and permissions handled for you and no platform engineering required. Its pricing page publishes no figure at all, and every route ends at contact sales. That is the trade: Glean is far less work and you will not know what it costs until you are in a procurement cycle.

Open WebUI is the most popular self-hosted AI workspace by a wide margin, at 150,623 stars and 22,000 forks, and it is free to run. It is a better chat interface than Cloudflare OS out of the box. It is not a governance layer: there is no Gatekeeper equivalent holding credentials away from the model, and GitHub’s licence detector does not return a standard OSI identifier for it, where Cloudflare OS is plainly Apache-2.0.

LibreChat is the cleaner-licensed alternative in the same space, MIT licensed with 42,698 stars and 8,835 forks, free to self-host, and happy to run against any provider. Like Open WebUI it gives you the workspace and leaves the hard part to you. Choose either of these if you want a chat surface for your company. Choose Cloudflare OS if what you actually need is the policy layer between the agents and your systems.

Final Verdict: Is Cloudflare OS Worth It?

Cloudflare OS is the most serious open source attempt yet at rolling AI out to a whole company rather than to its engineers. The Gatekeeper design is the right answer to the credential problem, personal apps on Durable Objects are a real architecture rather than a demo, and routing every inference call through AI Gateway gives you cost controls that most companies are currently improvising. Kenton Varda writing most of it and Cloudflare running it internally for thousands of people are both strong signals. It is free, and for a team under 50 already on Cloudflare it is close to free to operate too.

Two things keep it off a higher score. It is unfinished in the ways that matter for something company-wide: no tagged releases, 111 open issues, and a managed version still on the roadmap. And the pricing story is honest by omission rather than by disclosure. “Free and open source” is true of the software and misleading about the deployment, because Access at $7 per user per month over 50 people, Durable Objects usage, log retention, Guardrails inference and a partner engagement are all real costs nobody totals for you. If you have the platform team, deploy it now. If you do not, wait for the dashboard version and read this review again then.