Cloudflare OS: an open platform for agents, apps, and work
Cloudflare OS: A Universal Framework for Agents, Applications, and Organizational Productivity
Every organization is driven by a core mission—a fundamental reason for its existence.
This mission, along with the accompanying jargon, operational procedures, internal systems, and cultural standards, is passed down to the employees. These individuals then blend that corporate context with their personal expertise to drive the mission forward.
The output of this work varies wildly: it could be source code, strategic slide decks, internal documentation, professional relationships, or tangible physical results.
The "Feedback Loop" Gap
Some outputs are binary. For instance, code either executes successfully or it crashes. Because of this clear feedback loop, AI agents have spent the last few years becoming highly proficient at generating functional code for developers.
However, applying this same efficiency to the rest of a business is significantly more complex. To be effective, agents require:
- Deep Context: An understanding of the company's specific goals and language.
- System Access: The ability to interact with the actual tools employees use.
- Actionable Output: The capacity to turn that access and context into meaningful progress.
Cloudflare OS solves this by providing every employee with a personalized agent and a workspace tailored to their company's specific knowledge base and infrastructure.
From Internal Beta to Open Source
In May of this year, Cloudflare deployed the initial version of Cloudflare OS internally. Thousands of staff members—many of whom are not engineers—now use it daily to:
- Draft documents and presentations.
- Automate repetitive manual tasks.
- Develop lightweight apps for data visualization.
A critical component of this was the shared library of context and skills. By capturing the "best-known way" to perform recurring tasks as agent instructions, the organization ensures that when one person optimizes a process, the entire company inherits that improvement.
Evolution: What We Learned
The open-source version of Cloudflare OS is a direct result of the lessons learned during internal deployment (as detailed by CIO Sam Rhea).
| Feature | Version 1 (Internal) | Cloudflare OS (Open Source) |
|---|---|---|
| Focus | Individual private workspaces | Collaborative, company-wide ecosystem |
| App Nature | Static outputs | Live software connected to internal systems |
| Execution | Agent-driven (token heavy) | Deterministic workflows (efficient) |
| Tooling | Basic MCP server access | Resource-aware observation and access |
| Security | User-implemented | Platform-native governance |
Previously, security was a responsibility left to the individual app creator. Now, security is an intrinsic part of the platform foundation, ensuring that collaboration never accidentally leaks restricted data.
Introducing the Cloudflare OS Architecture
At first glance, Cloudflare OS looks like a standard browser-based AI chat. However, it is fundamentally different because every interaction is grounded in curated organizational skills and context.
The Three Core Pillars
The platform is composed of three integrated layers:
- The Agent Workspace: A hub for conversations grounded in company knowledge, featuring an isolated runtime where agents can write and execute code on the fly.
- Security & Governance Framework: A robust system ensuring that access to internal services and sensitive data is handled safely.
- Modifiable App Platform: A system for creating personal apps that can be shared, edited, and evolved by teammates.
Capabilities and Use Cases
The goal is to transform a simple conversation into a permanent asset. Whether it is a document, a functional app, or an automated workflow, the work persists.
🛠️ What you can achieve:
- Advanced Research: Instead of stuffing massive datasets into a context window, the agent writes code to information directly from company resources.
- Dynamic Content Creation: Generate slides or spreadsheets that stay linked to live data. These can be exported to services like Google Drive.
- Collaborative App Building: When a spreadsheet is insufficient, agents can build full apps with custom logic and state, allowing multiple users to collaborate.
- Deterministic Workflows: For tasks that follow a strict sequence with only minor points of human judgment, the workspace converts agent sessions into efficient, repeatable workflows.
Technical Implementation Example
To define a "skill" within the OS, the platform uses a structured approach to ensure the agent knows exactly how to interact with a system:
{
"skill_name": "Quarterly_Report_Generator",
"context_required": ["finance_api", "branding_guidelines"],
"runtime_permission": "isolated_sandbox",
"steps": [
"Fetch Q3 revenue from finance_api",
"Apply branding_guidelines to slide template",
"Summarize variances using LLM"
]
}
By shifting the burden of context from the user to the platform, Cloudflare OS ensures that employees no longer have to explain the same terminology or best practices to an AI every time they start a new session. The cost of operation is reduced as the system moves from stochastic agent guessing to deterministic execution: