Cloudflare Computer Introduces Virtual Filesystem for Durable Objects and AI Agents

Cloudflare has released open-source documentation and packages for Cloudflare Computer, a virtual filesystem that operates inside a Durable Object. According to the repository details, the system maintains authoritative state using SQLite and exposes execution surfaces via `workspace.runtime`.

Core Architecture and Backends

The repository outlines three primary execution backends that can be registered under stable IDs:

• Container Backend: Projects SQLite state into a sandbox container as a real FUSE mount using a daemon called `computerd`. This provides a Linux userland with real binaries and networking, though the documentation labels this specific component as experimental.
• Isolate Shell: Executes bash commands directly within a Dynamic Worker by reaching the authoritative workspace via Workers RPC, avoiding secondary stores or synchronization round-trips.
• Isolate JavaScript: Runs ECMAScript modules in a fresh Dynamic Worker featuring structured inputs and results, durable relative imports, and a workspace-backed Node file system implementation.

Developers can also construct a workspace without any backend, utilizing the filesystem standalone.

Package Status and Implementation

While the primary `@cloudflare/computer` package is noted as suitable for production use, various subcomponents—including the container backend, RPC helpers, and specific daemons—are explicitly marked as experimental or work-in-progress. The published specifications under the `docs/` directory represent forward-looking design intent rather than current code guarantees.

Practical Steps and Verification

To verify and build with the platform, developers can inspect the repository structure, install the `@cloudflare/computer` package, and review individual package README files.

• Installation: Install the package via npm and follow package-specific setup guidelines.
• Examples: The repository includes a diverse `examples/` directory containing runnable workspace scenarios, ranging from basic agent loops and container configurations to asset generation and runtime comparisons.
• Contribution: The project accepts bug reports, feature proposals, and feedback through GitHub issues and discussions, while unsolicited pull requests are not accepted. Collaborators must follow specific guidelines outlined in `CONTRIBUTING.md` and `COLLABORATORS.md`.

For exact performance benchmarks, developers can consult `docs/19_performance.md`, which details metadata versus sequential I/O performance characteristics.