Why AI tools know nothing about your company — until now
The New Stack Adrian Bridgwater ● Covered by 6 sources
Cloudflare just launched CloudflareOS, an open-source AI workspace that actually knows how your company works. It skips raw API keys for tighter, capability-based access, so agents can't just grab everything.
Based on reporting by The New Stack, Adrian Bridgwater — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Cloudflare rolled out CloudflareOS this week, and the pitch is blunt: most enterprise AI tools are brilliant about the world and clueless about your business. Every chat starts from zero, with staff re-explaining internal processes the AI should already know. CloudflareOS tries to fix that by baking a company's terminology, approval steps, templates and workflows into a workspace that travels with every employee from day one, rather than living in some analyst's head or a Slack thread nobody can find later.
What makes this more than a rebadged virtual desktop is the access model underneath it. Rita Koslov, Cloudflare's VP for developers and AI, told The New Stack that agentic work is pushing company data out of controlled systems in a way that never happened when people simply queried a data warehouse. Her answer is capability-based access: instead of handing an agent a broad API key, CloudflareOS grants access to one specific resource, logs exactly what the agent saw, and checks that anyone viewing the resulting work is actually allowed to see the source data too. It's a narrower door, deliberately.
Under the hood, Cloudflare leans on its own Dynamic Workers and Durable Objects Facets to run huge numbers of AI-generated apps — documents, spreadsheets, workflows, even full-stack tools — each in an isolated runtime with its own SQLite storage, outbound networking switched off by default, and Gatekeepers that expose only what's explicitly permitted. Koslov argues the hard problem was never generating an app; it's running thousands of them safely with persisted state, which is why these components had to be invented rather than bolted on.
Outside voices are more skeptical about where the real cost sits. Vladimir Beskorovainyi, an enterprise AI architect and founder of Besk Tech, told The New Stack that pricing AI platforms by the token misses the point entirely — what a company is actually buying is the discipline of writing down how its processes really work, and keeping that description accurate as the business changes. He likes that Cloudflare breaks spending down by person, team and app, calling it the first time he's seen a vendor treat cost as an engineering signal instead of just a bill. But he's clear that owning your context isn't the same as that context being correct, and that the work of keeping it true stays entirely with the customer.
CloudflareOS started as the internal platform Cloudflare built to run its own workforce, and thousands of its employees already use it for research, live-data documents and app-building. Now it's open source, running inside a company's own Cloudflare account and compatible with any AI model through Cloudflare's AI Gateway — so nobody's locked into a single provider, and routine tasks can be routed to cheaper models. Matthew Prince, Cloudflare's co-founder and CEO, framed the whole project simply: they built it because nothing else did what they needed.
My take — AI-written commentary, not fact-checked reporting
The access model here is the genuinely interesting part — swapping raw API keys for narrow, logged, capability-based grants is the kind of boring security fix enterprises should have demanded years ago. But Beskorovainyi's warning deserves more attention than it'll get: buying the platform is easy, keeping the context file honest as the business changes is the actual job, and most companies will underfund exactly that ongoing work while congratulating themselves on the shiny new workspace.
Read more about this at: The New Stack