cd ~/series/demystifying-ai
Cloudflare OS - The Open-Source AI Workspace With 6,775 GitHub Stars
EP 20 NEW RELEASE GitHub Aug 20, 2026 5 min

Cloudflare OS: The Open-Source AI Workspace With 6,775 GitHub Stars

Cloudflare open-sourced the AI workspace their own employees use daily. Thousands of people build documents, automate tasks, and ship apps on it. Ars Technica called it a "vibe-coding platform for people who aren't coders."

Share:
// TL;DR

Cloudflare open-sourced the AI workspace their employees use daily. 6,775 GitHub stars in 5 days. Browser-based agents grounded in company context. "Gatekeepers" control what agents can see, change, and when humans approve. Zero Trust by default. Every app runs in an isolated V8 runtime with its own SQLite database. Apache 2.0.

# Architecture

The entire system runs on Cloudflare Workers. No containers. No VMs. Every workspace is a Durable Object. Every app is a Dynamic Worker Facet with its own SQLite database.

BROWSER
Your team opens the workspace in any browser
AGENT WORKSPACE
Agents grounded in company context and skills
Durable Object State + Memory
GATEKEEPERS
Zero permissions by default. Per-task grants. Human approval gates.
INTERNAL SYSTEMS
Your data, APIs, and services
GADGETS (Apps)
Dynamic Worker Facets with own SQLite
KEY DETAIL

Every app gets isolated runtime via V8 isolates, not containers. Each workspace is its own Durable Object. Each Gadget is a Dynamic Worker Facet with its own SQLite database. Deterministic steps run first - AI inference fires only when needed, not as token-hungry sessions.

# The Gatekeeper Model

This is the core innovation. Gatekeepers are governed connectors that give system owners precise control over what AI can see, change, and when humans must approve.

Traditional AI Access
Full access to everything
Hope the model behaves
No human approval gates
System owner has no control
Cloudflare OS Gatekeepers
Zero permissions by default
Per-task access grants
Human approval gates built in
System owner controls everything
ZERO TRUST BY DEFAULT

Agents start with zero permissions. Access is granted per-task. System owners define what data the agent can read, what actions it can take, and which actions require human approval. This is not prompt-level guardrails - it is infrastructure-level access control.

# The Numbers

6,775
GitHub Stars
641
Forks
Apache 2.0
License
1000s
Internal Users
V8 Isolates
Runtime Isolation
0
Containers
Multi-Model Selection

Cloudflare AI Gateway routes inference. Not every task needs a frontier model. Admins control which models are available, attribute every inference call, and set budgets.

Two Repos

Core repo contains the full workspace platform. Starter repo is a deployment template. Requires Cloudflare Workers Paid plan.

# What You Can Build

Thousands of Cloudflare employees use this daily. These are not demo apps - they are production workflows running inside the company right now.

Document Generation

Create documents and slides grounded in company data. Agents pull context from internal systems through Gatekeepers.

Workflow Automation

Automate repetitive tasks with deterministic steps. AI inference fires only when needed, keeping costs predictable.

Data Visualization Apps

Build small apps (Gadgets) with their own isolated SQLite database. Each runs in its own V8 isolate with zero shared state.

Internal Tools

Deploy internal tools that connect to existing systems. Gatekeepers ensure agents only access what they are authorized for.

ARS TECHNICA

"A vibe-coding platform for people who aren't coders."

// Bottom Line

Cloudflare OS is not another chatbot wrapper. It is a production AI workspace with real security architecture. Gatekeepers give system owners precise control. Agents start with zero permissions. Every app gets isolated runtime. Thousands of Cloudflare employees use it daily, and now anyone can deploy it to their own infrastructure. The model choice is yours, the data stays yours, and the security model actually works.

NEXT EPISODE
Sat
#21 Upcoming

Mixture-of-Kittens: Cursor Open-Sourced Their MoE Training Kernel

Cursor open-sourced the MoE megakernel behind Composer. 2.37x faster than DeepSeek. Deterministic. Already training on tens of thousands of GPUs. Apache 2.0.

Enjoyed this?

New episodes Mon, Wed, Sat.