Cloudflare Computer

Cloudflare Computer is an open-source agent runtime introduced by Cloudflare in August 2026 that gives every AI agent its own "computer". It provides a persistent, SQLite-backed virtual filesystem living inside a Durable Object, plus pluggable execution backends — so agents can read, write, edit files, run shell commands, work with Git, and publish projects without any VPS, Docker daemon, or container orchestration of your own.

Key Features

• Durable Virtual Filesystem: A workspace backed by SQLite in a Durable Object that survives restarts, with R2-backed mounts for pre-filling read-only data.
• Three Execution Backends: A full Linux container (real binaries, real network, FUSE-mounted workspace synced via capnweb), a lightning-fast just-bash isolate shell in a Dynamic Worker, and an isolated JavaScript (ECMAScript module) runtime.
• Runtime Selection Built In: The workspace.runtime.exec() entry point lets the model pick the right backend per task — fast cheap isolates for file/data work, containers only when real userland is needed.
• AI SDK Toolkit: Ready-made tools (read, write, edit, ls, exec, publish) compatible with the AI SDK, usable with @cloudflare/agents and other agent frameworks.
• Git & Artifacts Support: Built-in git client (ws:git) and Cloudflare Artifacts integration (ws:artifacts) for cloning, committing, and publishing projects.
• Gated, Audited, Observed: All file and execution operations are permission-gated and logged, giving a clear paper trail of what the agent did.

Use Cases

Who Should Use This Tool?

• Agent Platform Builders: Teams shipping autonomous agents that need reliable, durable file state and code execution without managing VMs.
• Edge Developers: Anyone already on Cloudflare Workers, Durable Objects, and R2 who wants code execution co-located with their stack.
• Indie Hackers: Solo developers who want a $0 option to give their agent a real Linux environment, npm, Node, and Git.

Problems It Solves

1. State loss across agent runs: The SQLite-backed workspace survives restarts, so long-running agents keep their files.
2. Execution environment sprawl: One runtime dynamically routes between isolates and containers instead of forcing a single primitive.
3. Infrastructure cost: No VPS, no Docker, no API bill — an agent's computer runs inside Workers with free-tier-friendly usage.

Pricing

Advantages & Unique Selling Points

Compared to Competitors:

1. No container by default: Most sandboxes boot a VM or container per task; Cloudflare Computer runs most work in isolates and falls back to containers for under ~10% of tasks.
2. Deep edge integration: Native Durable Objects persistence, R2 mounts, and Workers bindings that standalone sandbox providers can't match.
3. Open source: The full @cloudflare/computer package is on GitHub with examples, unlike closed managed sandboxes.

What Makes It Stand Out:

• The "filesystem as a durable object" model — state and execution live in the same place.
• AI SDK-compatible tooling designed for frontier models to route themselves.
• Free to start, with a genuinely useful free tier.

User Reviews

"Giving agents a persistent filesystem in a Durable Object is the right abstraction. No VPS, no Docker, no API bill." — Social post shared after the launch announcement

"The runtime selection works — frontier models are surprisingly good at choosing the fast isolate for file work and containers only when they need real binaries." — Commenter on the Cloudflare blog

Getting Started

Quick Start Guide

1. Install the package: npm install @cloudflare/computer
2. Create a Workspace: Instantiate a Workspace on any Durable Object with { storage: this.ctx.storage }.
3. Attach backends: Add a CloudflareContainerBackend, WorkerShellBackend, or WorkerJavaScriptBackend.
4. Hand tools to your agent: Use @cloudflare/computer/tools to expose read/write/edit/ls/exec to any AI SDK agent.

Integration

Integrates seamlessly with:

• Cloudflare Workers, Durable Objects, R2, and Cloudflare Containers
• AI SDK and @cloudflare/agents
• any-isomorphic-git based workflows

Frequently Asked Questions

Does it run Linux binaries?

Yes, the container backend projects the SQLite workspace into a sandbox container as a real FUSE mount with a full Linux userland, real binaries, and real network.

Is the filesystem really persistent?

Yes, the Durable Object holds the authoritative state in SQLite, so all file operations survive DO restarts.

Is it free?

The library is open source and self-hostable; you pay only for Cloudflare usage underneath. There is no separate per-seat agent fee.

Alternatives

If Cloudflare Computer isn't the right fit, consider these alternatives:

• E2B: Best for Firecracker-microVM isolation on a dedicated sandbox platform.
• Daytona: Better if you need persistent full dev environments with GPU sandboxes.
• Modal: More suitable for Python/ML workloads with GPU-backed compute.

Tips & Best Practices

1. Let the model route: Keep both isolate and container backends available and rely on tool descriptions to pick the right one.
2. Use R2 mounts for read-only data: Pre-fill large datasets without copying them into the workspace.
3. Audit everything: Keep gating and logging enabled so long-running autonomous agents leave a recoverable paper trail.

Conclusion

Cloudflare Computer turns Cloudflare's edge into a cheap, durable "computer" for AI agents — a persistent filesystem plus switchable execution environments, all open source and free to start. If you're building agents on Workers or want to stop paying for idle VPS capacity, it's worth a serious look.