Cloudflare Computer

There's a pattern behind the best AI agents right now: give them a computer. Coding agents already work this way, poking around a filesystem, running shell commands, testing changes, and iterating until something works. Cloudflare wants to formalize that pattern into a product, and this week it shipped an early preview of a package called @cloudflare/computer, meant to be the plumbing underneath any agent that needs to act on files and run code.

The backstory matters here. At the start of this year, spinning up a container per agent was the default approach. Over the last few months, agent frameworks have shifted toward separating the "brain" (the model doing the reasoning) from the "hands" (a sandboxed environment doing the actual work). Cloudflare argues that approach still hits a wall: there simply isn't enough compute across every cloud provider combined to hand every single agent its own containerized environment once you're talking about hundreds of millions or billions of concurrent agents. That's the panic behind the current scramble for CPU compute, not just GPUs.

Cloudflare's answer traces back almost a decade, to the bet it made with Workers, and again nearly six years ago with Durable Objects. Both rely on isolates rather than full containers, and isolates scale differently: they spin up and shut down fast, they can hibernate when an agent goes idle, and they can hold onto an agent's state without needing a dedicated machine sitting around. Last year Cloudflare let isolates call out to container sandboxes on demand as a tool, so heavier compute only gets used when a task actually needs it.

@cloudflare/computer tries to collapse that combination into something simpler. Instead of developers manually wiring isolates and containers together, the package hands an agent a single, declaratively defined filesystem plus a choice of execution environments, each with its own tradeoffs on speed, capability, and cost. File manipulation or a git operation can run inside a cheap isolate; anything needing a real Linux environment, npm, or a native binary falls back to a container. Both operate against the same synced files, and every operation is gated, audited, and logged so there's a record of what the agent actually did.

Under the hood, the workspace is a virtual filesystem backed by SQLite. The isolate runtime translates shell commands into JavaScript using something called just-bash, running inside a dynamic worker with direct bindings to the filesystem. The container runtime mounts the same filesystem via FUSE so a full Linux environment can read and write to it too. Both plug into an AI SDK-compatible toolkit with familiar primitives — read, write, edit, ls, and a shell exec tool that picks the right backend based on the task, something Cloudflare says frontier models are already quite good at doing on their own.

Cloudflare is already running this internally for jobs like building and deploying JavaScript apps, generating customer documentation, and browsing the web, and its stated target is for containers to handle less than 10% of an agent's actual workload, with isolates doing the rest. The package is open-source and available now via npm, and Cloudflare is explicitly framing this as an experiment it wants to iterate on with developers pushing agents at scale, rather than a finished product.