Today we are releasing Cloudflare OS, a chatbot with connectors, just like every other tech company is doing.
Except actually, it's different. This is a remake of Sandstorm[.]io, my startup from 10 years ago, except this time built on Cloudflare Workers (the platform I've spent the last 9 years building) and deeply leveraging AI. This is more or less the culmination of my secret 10-year master plan.
This is a full-on personal app vibe coding platform, in which the sandbox is so secure that you can pretty much go wild -- the AI cannot introduce a significant security bug. We believe a company's security team can feel comfortable giving non-technical users permission to vibe code and then sleep soundly at night.
How is that possible? It's the Sandstorm security model, revisited. A "Gadget" is the same thing as a Sandstorm "Grain": a fine-grained app instance. For example, if you have a document editor app, each document runs as a separate instance of the app, in a separate sandbox (one "Gadget").
This means two things, both of which I think are Big Deals:
1. The platform can manage all access control, by controlling who can access the Gadget at all. There is no way the Gadget can accidentally leak itself to an attacker -- even an attacker who has access to other Gadgets based on the same app.
2. Since everyone is running their own copy of the code, everyone can freely modify their copy of the code.
Think about #2 a bit more.
What if, when you wanted a new feature in the software you are using, you could just prompt your agent to add it?
This doesn't work in the cloud Software-as-a-Service model, because you are not running your own copy of the app.
Sandstorm tried to change that 10 years ago, but the world wasn't ready, because not enough people had the skills or patience to actually modify their software. AI has changed that. Now you just ask the agent -- the same agent that you are using to help you interact with the Gadget can also modify the code of the Gadget.
The article posted buries the lead. Cloudflare OS reads like almost any other AI knowledge base until midway when apps are introduced and then becomes overly technical with a code snippet.
Of course, personal apps are more useful if they can connect to external services. Cloudflare OS introduces a "connector" system we call Gatekeepers. This is sort of like MCP (and MCP is supported as a kind of Gatekeeper), but with a lot more:
* Instead of exposing tools, a Gatekeeper exposes a Cap'n Web RPC API. That makes it appropriate for use by both agents (via code mode) and Gadgets.
* Gatekeepers integrate with the Cloudflare OS UI to provide inline audit logging and human-in-the-loop approvals for all side-effecting actions.
* When an action requires approval, the agent does not need to stop and wait for it. A Gatekeeper will simulate the outcome, allowing the agent to keep running and queue up more work. You can then approve everything in a batch at the end. Hopefully, this means you no longer feel the need to turn on auto-approve! (But you still can if you want.)
We have already built Gatekeepers for a huge number of services, from GitHub to Home Assistant. We've found, with the right skills, AI can basically crank these things out for any given API, solving the chicken-and-egg ecosystem problem. We are working on a "software factory" of sorts where you can input the API docs for your favorite service and get a well-written Gatekeeper. That's not quite ready yet -- so for now use the ones we provide or use the write-gatekeeper skill to write your own. You can add your own Gatekeepers to your personal deployment.
> * When an action requires approval, the agent does not need to stop and wait for it. A Gatekeeper will simulate the outcome, allowing the agent to keep running and queue up more work. You can then approve everything in a batch at the end. Hopefully, this means you no longer feel the need to turn on auto-approve! (But you still can if you want.)
Can you say more about how you simulate outcomes of approvals? For example, if I need an approval to be granted permission to read a document, how could it simulate the document contents?
When you share a Gadget, we verify that anyone you share with also has direct permission to access each of the resources it is connected to (via the Gatekeeper system). Hence, no security bug in the Gadget itself could accidentally grant people access to things they don't already have.
Everytime I read about new things from Cloudflare they look really cool but I cannot shake the feeling of not wanting to use them for fear of lock-in, am I too paranoid?
Yes. Be locked in. It's better to go 100% and only pay the switch costs if you have to.
Developing abstractions to let you switch technologies has 2 costs: The actual development, and not fully utilizing a specific technology. Lots of people prefer pain in small doses because it can be rationalized as eventual payoff.
There are degrees of this decision, of course, but I would say most folks worry too much about this particular champagne problem.
Cloudflare seems mature enough at this point that if you're willing to accept any kind of vendor lock-in, then accepting it through them is likely acceptable. I won't say they're on-par with the big cloud providers, but if they offer the services you need and you'd be willing to lock-in with AWS, etc., then they should be considered.
If you're generally opposed to any vendor lock-in, then Cloudflare is definitely a hard sell. Many of their services are pretty specific to their ecosystem and I imagine trying to keep things portable would be infeasible any many scenarios. Granted, this is true in a provider like AWS, too, if you were depending heavily on their proprietary, managed services. But there's definitely value in knowing that you can stick to "generic" services as much as possible in AWS and be able to move away relatively easily.
Cloudflare is definitely becoming a "real" cloud provider (I'd say their release of R2, for me, pushed them into the "serious contender" category), but they focus on pretty specific features that aren't very portable. This Cloudflare OS service is a good example of that. If it is exactly what you need, then it is worth checking out and taking on the risk of vendor lock-in. But if you're doing something where long-term planning and portability is important, then you'll have a hard time warranting even considering them.
I feel exactly the same. The CF services I do use are DNS and R2. I'd be curious if there's a guide somewhere that lists their services through the lens of lock-in. For example, R2 is somewhat S3 compatible so it's less risky to use because I can swap providers.
Yeah, with love the words "open source" mean a lot less here (de facto, of course not de jure) when the resulting open source product is entirely contingent on Cloudflare's PaaS.
And I think they suffer a little a little bit from their infrastructure being _comparatively_ unique. As providers go, they're on the harder end to replace with something else if you were to ever want to retarget your workloads.
Cloudflare is just another corporation with greed in their eyes and if you don't design yourself to their puzzle than you're not welcomed.
So, no, you're not. This is another stab at gatekeeping the upcoming Ai era of the internet. It's no different to the playbook of Google, Apple, and Facebook. An extra wall in walled garden land. One you will need to jump if you want to break free.
In terms of security: Firefox on FreeBSD instantly triggers loops of "prove your human" on their turnstile.
I can't use old versions of android without being denied and VPNs triggers all. Anything that they don't see as "secure" you're denied from. Vendor lock-in at it's best and Google & Co, would be jealous if they were not in bed with.
This is "Your LLM/Ai isn't using our service, so your Ai isn't allowed within our garden" netting and what does Cloudflare offer as a service? A turnstile to the garden of things.
Remember folks, no corporation gives a rat's ass about you. Internally, or externally, you're just another sheep to extract wool from.
"OS for work" - like, what does it mean? OS is operating system, it runs on a hardware and provides you to operate the system. I hate when companies destroy the English language with vibe invented terms.
This is not a traditional computer operating system. We use the term "operating system" in two senses:
- An operating system for the company to be productive with AI, in a way that is safe, so that the security team can sleep at night.
- An operating system for AI workloads, analogous to the sense in which a traditional operating system manages compute workloads.
Because AI enables a new layer more akin to an actual OS. But also because I imagine they all want all the things running through them and an OS is the widest net they can cast in that direction.
I think because the idea of an "agent operating system" is starting to feel more and more inevitable. Since I can use AI to:
- Pull information from external sources
- Transform and/or augment that information
- Publish information
I almost do not need a traditional OS anymore. This obviously isn't entirely true: I still use it to interact directly with humans, and it's the way I access my LLM. For now.
This represents a small part of what IT is worried about. Offer a beer and say the word "SharePoint" if you want the long version of where this could wind up.
The issue isn't letting end-users add features. The issue is twelve end-users customizing something that produces output, saving it, and now nobody can read it.
This can get solved. If it even gets a chance to gain traction (it should, though IT will of course push back). But let's not pretend we're there yet.
This is effectively a Codex/Claude app competitor. Good move from Cloudflare since it helps them sell their core infra offerings. I like the direction they’re going with access control.
The real question is, can Cloudflare become a equal or better OpenAI / Claude provider, while offering a vastly more robust platform than either?
I think the answer is yes, simply because both OAI and Anthropic leverage CF for their hosting today.
I used to meme back in 2008 (when Google docs first was becoming big) that "The problem has always been the sandbox". The problem Chrome and Firefox solved as the Sandbox - you can browse some random code across the web without worrying about getting hacked. The same problem has arisen today but now with AI agents.
There is likely to be huge winners in the AI space simply based on who can create the best way to contain AI agents and "expose AI value" safely.
So, that means even more pressure on OpenAI and Anthropic. I really don’t see how they will stay competitive long term. Cloudflare can offers way better product by integrating with the rest of their platform and offering a variety of models, including open models they host themselves
I can't help but wonder if these things that are now emerging have a sort of "borrowed time" component tied to them.
I see the examples, and they clearly tackle both organizational inefficiencies and human connection/communication as the core problems.
Sounds like things are converging to "my AI will talk to your AI" and the role of humans will just be a human "router" in the loop switching up nodes in a graph, like the return of Switchboard Operators from the 1950s.
This is not a traditional computer operating system. We use the term "operating system" in two senses:
An operating system for the company to be productive with AI, in a way that is safe, so that the security team can sleep at night.
An operating system for AI workloads, analogous to the sense in which a traditional operating system manages compute workloads.
Engineers should be wary about being forced into using particular prompts/skills/instructions.
Why would I use a shared LLM-in-the-cloud, instead of my own? The shared one precludes any possibility to customize or differentiate my workflow, prompts, or skills. As an employee that sounds like a death sentence to doing my job effectively.
> Cloudflare OS is available today and is open source. Check out the cloudflare-os GitHub repository. You can deploy it into your own Cloudflare account and use your own Access policies, AI Gateway configuration, data, and integrations.
It's open source, but it is so incredibly tied to their platform, that there is no vendor portability, which has made almost every of their product launches post-Workers kind of meh.
Everyone is building these same looking apps that look like Codex, Claude desktop. We don't need these. Slackbots work more than well enough to collaborate with agents.
This is the middle of Cloudflare's "Agents Week"... and I believe they had an Agents Week last year. We're like... at least a year or two into Cloudflare being in the AI party. ;)
Text from tweet:
Today we are releasing Cloudflare OS, a chatbot with connectors, just like every other tech company is doing.
Except actually, it's different. This is a remake of Sandstorm[.]io, my startup from 10 years ago, except this time built on Cloudflare Workers (the platform I've spent the last 9 years building) and deeply leveraging AI. This is more or less the culmination of my secret 10-year master plan.
This is a full-on personal app vibe coding platform, in which the sandbox is so secure that you can pretty much go wild -- the AI cannot introduce a significant security bug. We believe a company's security team can feel comfortable giving non-technical users permission to vibe code and then sleep soundly at night.
How is that possible? It's the Sandstorm security model, revisited. A "Gadget" is the same thing as a Sandstorm "Grain": a fine-grained app instance. For example, if you have a document editor app, each document runs as a separate instance of the app, in a separate sandbox (one "Gadget").
This means two things, both of which I think are Big Deals: 1. The platform can manage all access control, by controlling who can access the Gadget at all. There is no way the Gadget can accidentally leak itself to an attacker -- even an attacker who has access to other Gadgets based on the same app. 2. Since everyone is running their own copy of the code, everyone can freely modify their copy of the code.
Think about #2 a bit more.
What if, when you wanted a new feature in the software you are using, you could just prompt your agent to add it?
This doesn't work in the cloud Software-as-a-Service model, because you are not running your own copy of the app.
Sandstorm tried to change that 10 years ago, but the world wasn't ready, because not enough people had the skills or patience to actually modify their software. AI has changed that. Now you just ask the agent -- the same agent that you are using to help you interact with the Gadget can also modify the code of the Gadget.
And it is so fun.
Buzz looks like a more reasonable direction than this does.
https://github.com/cloudflare/cloudflare-os
It runs on the open source Workers Runtime:
https://github.com/cloudflare/workerd
You can run it at home if you want. There's even a Home Assistant Gatekeeper.
And it would be functionally the same? With the same performance if you put enough hardware there?
If so cool, and that should have been clearly in the announcement, because as it sounded it is “I am a sucker, lock me in!”
Yes. It even supports ollama and works decently well with some local LLMs.
> And it would be functionally the same? With the same performance if you put enough hardware there?
Honestly it is faster running locally.
> If so cool, and that should have been clearly in the announcement,
We had so many things we wanted to say, it was hard to fit them all into one story.
The blog post is for an enterprise audience. My tweet thread is for the hacker audience.
The article posted buries the lead. Cloudflare OS reads like almost any other AI knowledge base until midway when apps are introduced and then becomes overly technical with a code snippet.
What gets buried is the "lede". It is a really unusual thing — a deliberate, engineered misspelling of "lead" that leaked into broader use:
https://www.merriam-webster.com/wordplay/bury-the-lede-versu...
My tweet thread, and the GitHub readme, is really the presentation intended for the HN crowd. The blog post is for a different audience.
https://github.com/cloudflare/cloudflare-os
Yep, that's precisely the vision I formulated some months ago: https://manuel.kiessling.net/2025/11/04/what-if-software-shi...
This can only be correct when the application can’t affect anything outside the sandbox. Which would significantly restrict useful applications.
https://x.com/KentonVarda/status/2084990302784307363
-------------
Of course, personal apps are more useful if they can connect to external services. Cloudflare OS introduces a "connector" system we call Gatekeepers. This is sort of like MCP (and MCP is supported as a kind of Gatekeeper), but with a lot more:
* Instead of exposing tools, a Gatekeeper exposes a Cap'n Web RPC API. That makes it appropriate for use by both agents (via code mode) and Gadgets.
* Gatekeepers integrate with the Cloudflare OS UI to provide inline audit logging and human-in-the-loop approvals for all side-effecting actions.
* When an action requires approval, the agent does not need to stop and wait for it. A Gatekeeper will simulate the outcome, allowing the agent to keep running and queue up more work. You can then approve everything in a batch at the end. Hopefully, this means you no longer feel the need to turn on auto-approve! (But you still can if you want.)
We have already built Gatekeepers for a huge number of services, from GitHub to Home Assistant. We've found, with the right skills, AI can basically crank these things out for any given API, solving the chicken-and-egg ecosystem problem. We are working on a "software factory" of sorts where you can input the API docs for your favorite service and get a well-written Gatekeeper. That's not quite ready yet -- so for now use the ones we provide or use the write-gatekeeper skill to write your own. You can add your own Gatekeepers to your personal deployment.
Can you say more about how you simulate outcomes of approvals? For example, if I need an approval to be granted permission to read a document, how could it simulate the document contents?
When you share a Gadget, we verify that anyone you share with also has direct permission to access each of the resources it is connected to (via the Gatekeeper system). Hence, no security bug in the Gadget itself could accidentally grant people access to things they don't already have.
What particular feature/capability made you think this?
I do not disagree but i am curios. Lets have a discussion instead of just throwing feel good vibes around...
https://github.com/cloudflare/cloudflare-os
It runs on top of our open source runtime.
https://github.com/cloudflare/workerd
Developing abstractions to let you switch technologies has 2 costs: The actual development, and not fully utilizing a specific technology. Lots of people prefer pain in small doses because it can be rationalized as eventual payoff.
There are degrees of this decision, of course, but I would say most folks worry too much about this particular champagne problem.
If you're generally opposed to any vendor lock-in, then Cloudflare is definitely a hard sell. Many of their services are pretty specific to their ecosystem and I imagine trying to keep things portable would be infeasible any many scenarios. Granted, this is true in a provider like AWS, too, if you were depending heavily on their proprietary, managed services. But there's definitely value in knowing that you can stick to "generic" services as much as possible in AWS and be able to move away relatively easily.
Cloudflare is definitely becoming a "real" cloud provider (I'd say their release of R2, for me, pushed them into the "serious contender" category), but they focus on pretty specific features that aren't very portable. This Cloudflare OS service is a good example of that. If it is exactly what you need, then it is worth checking out and taking on the risk of vendor lock-in. But if you're doing something where long-term planning and portability is important, then you'll have a hard time warranting even considering them.
Also check out Bunny.net.
And I think they suffer a little a little bit from their infrastructure being _comparatively_ unique. As providers go, they're on the harder end to replace with something else if you were to ever want to retarget your workloads.
Not remotely paranoid.
So, no, you're not. This is another stab at gatekeeping the upcoming Ai era of the internet. It's no different to the playbook of Google, Apple, and Facebook. An extra wall in walled garden land. One you will need to jump if you want to break free.
In terms of security: Firefox on FreeBSD instantly triggers loops of "prove your human" on their turnstile.
I can't use old versions of android without being denied and VPNs triggers all. Anything that they don't see as "secure" you're denied from. Vendor lock-in at it's best and Google & Co, would be jealous if they were not in bed with.
This is "Your LLM/Ai isn't using our service, so your Ai isn't allowed within our garden" netting and what does Cloudflare offer as a service? A turnstile to the garden of things.
Remember folks, no corporation gives a rat's ass about you. Internally, or externally, you're just another sheep to extract wool from.
The issue isn't letting end-users add features. The issue is twelve end-users customizing something that produces output, saving it, and now nobody can read it.
This can get solved. If it even gets a chance to gain traction (it should, though IT will of course push back). But let's not pretend we're there yet.
"Your Cloudflare account is on the Free plan, but the Cloudflare OS backend needs Dynamic Workers, which requires the Workers Paid plan."
I think the answer is yes, simply because both OAI and Anthropic leverage CF for their hosting today.
I used to meme back in 2008 (when Google docs first was becoming big) that "The problem has always been the sandbox". The problem Chrome and Firefox solved as the Sandbox - you can browse some random code across the web without worrying about getting hacked. The same problem has arisen today but now with AI agents.
There is likely to be huge winners in the AI space simply based on who can create the best way to contain AI agents and "expose AI value" safely.
I see the examples, and they clearly tackle both organizational inefficiencies and human connection/communication as the core problems.
Sounds like things are converging to "my AI will talk to your AI" and the role of humans will just be a human "router" in the loop switching up nodes in a graph, like the return of Switchboard Operators from the 1950s.
This is not a traditional computer operating system. We use the term "operating system" in two senses:
An operating system for the company to be productive with AI, in a way that is safe, so that the security team can sleep at night. An operating system for AI workloads, analogous to the sense in which a traditional operating system manages compute workloads.
FTF[them]
Why would I use a shared LLM-in-the-cloud, instead of my own? The shared one precludes any possibility to customize or differentiate my workflow, prompts, or skills. As an employee that sounds like a death sentence to doing my job effectively.
https://github.com/cloudflare/cloudflare-os
You're supposed to customize it.
Also it features a whole system for adding your own skills and context.
In case anybody missed it.
EDIT: It let me down... (Something unexpected went wrong on our side.)
Please let me now if deploying it on your custom domain worked for you please...
> "Give every person an agent and workspace built around how your company works, what it knows, and the systems it relies on."
Hundreds of thousands of so called "AI startups" have been eliminated.
https://github.com/open-webui/open-webui