Cloudflare Computer: A Virtual Filesystem for Persistent AI Agents

Key points

• Stateful agents via filesystem: Cloudflare Computer stores all agent working state inside a Durable Object, using SQLite as the source of truth and exposing it as a virtual filesystem. Agents persist context by writing files, eliminating the need for vector databases or RAG pipelines.
• Filesystem mirrors developer workflow: Code, configs, logs, TODOs, and intermediate artifacts are already the substrate of human programming. By giving agents the same substrate, Cloudflare reuses decades of Unix tooling (grep, find, make, git) instead of inventing AI-specific memory abstractions.
• Decoupled execution, three pluggable backends:
• Container backend: Mounts the SQLite-backed state via FUSE into a real Linux container for full binary, network, and userspace access.
• Isolate shell backend: Runs just-bash (a pure-JS bash) inside a Cloudflare Worker, accessing state via RPC for fast, low-overhead execution.
• Isolate JavaScript backend: Executes an ECMAScript module in a Dynamic Worker with structured I/O, a node:fs/promises polyfill, and trusted ws:git / ws:artifacts modules.
A single Workspace can register multiple backends; workspace.runtime.exec(source, { backend }) is the unified entry point.
• Shared protocol for multi-agent collaboration: Files double as a collaboration medium. One agent's code, logs, or data become directly readable by another agent, removing the need for bespoke messaging protocols.
• Edge-native persistence: Durable Objects run across 300+ Cloudflare edge locations with automatic persistence and low latency, so agents remain warm and quickly wakeable rather than sitting idle on GPU clusters.
• Preview status: The README flags the project as PREVIEW ONLY — the API is unstable and not production-ready. Much of the documentation describes intended design rather than current implementation.
• Layered agent stack emerging: The post frames a four-layer architecture now developing independently:
1. Runtime layer — Codex, Claude Code, Cursor. 2. Loop control layer — projects like loopx. 3. State layer — Cloudflare Computer. 4. Skills layer — repositories like addyosmani/agent-skills. Filesystem is positioned as the common protocol tying these layers together.

Project facts

• Repository: github.com/cloudflare/computer
• Language: TypeScript
• Momentum: 796 stars in one day at time of writing
• Related Cloudflare project: github.com/cloudflare/agents (agent deployment framework)
• Related loop-control project: github.com/huangruiteng/loopx