Cloudflare has open-sourced Computer, a persistent virtual filesystem that gives AI agents a workspace that survives beyond a single session. Built on SQLite inside Durable Objects and running on Cloudflare's edge infrastructure, the preview release offers three execution backends — container, shell, and JavaScript — for developers to test agent workspaces today.
Cloudflare released Computer on June 5, 2026 as a public GitHub repository under the MIT license. The project ships as "give your agent a computer" — a virtual filesystem backed by SQLite inside a Durable Object, paired with a single execution entrypoint that lets AI agents read, write, and run commands in a persistent workspace.
The repo has accumulated 5,546 stars and 281 forks since launch. Two distinct execution models are available: an isolated JavaScript runtime for pure script execution, and a shell environment for command-line workflows. A container backend using FUSE mounting is also in development.
Cloudflare explicitly marks the project as preview-only, noting it is "not suitable for production use." The repository contains eight example projects spanning container setups, worker shells, JavaScript workers, and a compare-runtimes benchmark.
Key Takeaways
- Computer is an open-source virtual filesystem for AI agents, hosted at github.com/cloudflare/computer
- Built on SQLite inside Cloudflare Durable Objects for persistent state
- Three backends: Container (FUSE), Isolate shell (bash), Isolate JavaScript (ESM)
- Status is preview only; Cloudflare warns against production use
- Distinct from Cloudflare Kitesurf; Computer focuses on filesystem and execution, not browser automation
- 5,546 GitHub stars since creation on June 5, 2026
How Computer Works
Computer presents a virtual filesystem to AI agents through a Durable Object that stores all state in SQLite. The filesystem persists across agent invocations, meaning a workspace can span multiple sessions without losing files, state, or context.
The core API surface is a single function:
workspace.runtime.exec(source, { backend })Agents pass source code and select a backend. The three supported backends determine how the code runs:
-
Container: Mounts the filesystem via FUSE using the
computerddaemon with capnweb RPC -
Isolate shell: Runs commands through
just-bashinside a Dynamic Worker - Isolate JavaScript: Executes ECMAScript modules directly in a Dynamic Worker
This design lets agents operate in familiar environments — a shell for CLI workflows, JavaScript for programmatic tasks, or a containerized setup for more complex workloads.
Package Structure
The repository organizes code into four main packages:
- dofs: The Durable Object filesystem implementation
- rpc: Cap'n Proto RPC definitions for cross-boundary communication
- computerd: The FUSE daemon for container backends
- computer: The SDK and agent-facing API
Example projects ship alongside the core packages, demonstrating container setups, worker shells, JavaScript workers, a compare-runtimes benchmark, and an assets folder.
Preview Status and Limitations
Cloudflare's README states Computer is "in preview" and "not suitable for production." The project is actively developed — the last push was August 7, 2026 — but features and APIs may change.
The preview label means developers should expect breaking changes, incomplete documentation, and unpolished edge cases. The eight shipped examples are proof-of-concept implementations rather than production templates.
Computer vs. Kitesurf
Cloudflare runs multiple agent-facing projects. Computer is often mentioned alongside Kitesurf, but they serve different purposes.
Kitesurf is a browser automation tool that lets agents interact with web interfaces. Computer is a filesystem and execution surface — it gives agents persistent storage and a place to run code, not a browser to control.
The two projects could complement each other in a future architecture where an agent uses Kitesurf for browser tasks and Computer for file operations and script execution, but Cloudflare has not announced any integration between them.
For context on how agent products are advancing in both preview and production phases, see OpenAI Shuts Down ChatGPT Atlas on August 9.
Getting Started
Developers can clone the repository and run the example projects to test Computer locally. The tutorial example walks through basic setup, while the container and worker examples demonstrate the two active backends.
The MIT license allows commercial use, modification, and distribution without restriction.
Conclusion
Computer gives AI agents a persistent workspace — something many agent frameworks still lack. The SQLite-backed Durable Object approach means state survives between sessions without requiring external databases or file sync. The three-backend design covers shell, JavaScript, and containerized execution, though only the first two are production-ready at this stage.
The preview status is the main caveat. Cloudflare is clearly still shaping the project, and developers should treat it as a testing ground rather than a shipped product. For teams exploring agent workspaces on Cloudflare's edge, Computer is a meaningful step toward persistent, stateful AI systems.
Frequently Asked Questions
Q: Is Cloudflare Computer ready for production?
No. Cloudflare explicitly states the project is in preview and not suitable for production use.
Q: What databases does Computer use for persistence?
Computer uses SQLite inside Cloudflare Durable Objects. All filesystem state lives in a single SQLite database managed by the Durable Object.
Q: Can I use Computer with any AI framework?
Computer exposes a runtime.exec() API that any agent framework can call. It is not tied to a specific AI vendor or SDK.
Q: How is Computer different from Cloudflare Kitesurf?
Computer provides persistent filesystem storage and code execution. Kitesurf provides browser automation. They address different layers of agent infrastructure.
Q: What license covers the Computer project?
The repository is released under the MIT license, allowing free use, modification, and distribution.
Top comments (0)