Cloudflare OS: uma plataforma aberta para agentes, aplicativos e trabalho

Every company possesses a purpose, the driving motive behind its existence. Companies convey this purpose, along with their jargon, protocols, systems, guidelines, and customary work methods, to their employees. Individuals, armed with this context and their own expertise, strive towards the company's objective. Tasks can manifest in various forms, from programming to documents, presentations, collaborations, and tangible results in the real world.

Some are uncomplicated, with code either functioning correctly or not. Agents have been utilizing this iterative process to generate functional code for developers over the past couple of years. However, extending this approach to the broader organization poses a more formidable challenge. Agents must comprehend the organizational context and interact with the systems employed for daily operations.

They must translate this contextual knowledge and system access into productive work that advances the company's mission. This is the premise of Cloudflare OS. It provides every individual with an agent and workspace tailored to their organization: its operational dynamics, accumulated knowledge, and reliant systems. In May, we granted all Cloudflare employees access to the initial iteration of Cloudflare OS.

Thousands of individuals from diverse departments, many from non-engineering backgrounds, utilize it daily to compose documents, create slides, automate routine tasks, and construct small applications to visualize data and aid their work. Cloudflare OS also introduces a shared repository of contextual information and expertise cultivated by Cloudflare teams.

It encapsulates our terminology, procedures, and recommended practices for recurring work as instructions that agents can follow. When a team member discovers a more efficient approach, the entire organization benefits. Today, we are releasing an open-source version of Cloudflare OS. Any organization can deploy, integrate with internal systems, and adapt it to their specific needs.

The open-source Cloudflare OS is predicated on insights gained from the initial internal deployment, as detailed in our CIO, Sam Rhea's blog post. The first version centered on individuals collaborating through private workspaces. Apps were static rather than dynamic software tied to internal systems, and primarily deterministic jobs still necessitated running an agent skill multiple times, consuming additional model tokens.

Collaboration unveiled a more fundamental issue. While an MCP server provided insight into the tools an agent could invoke, it did not reveal the underlying resources observed by the agent. As individuals started sharing workspaces, apps, and outputs, it became imperative to ensure that collaboration did not inadvertently expose sensitive information that users were not authorized to view.

We have reconstructed Cloudflare OS on a novel foundation to address these challenges. Security must be an integral component of the platform, not an afterthought that each individual must implement independently. The outcome is a platform designed to serve the organization that deploys it. You can personalize the interfaces, integrate your tools, and incorporate the skills and context that reflect your organization's operations.

Cloudflare OS commences with a conversation in your browser, similar to other AI tools. What sets it apart is that each conversation is anchored in the context and skills your organization has curated. By setting a goal for your workspace, it can leverage this knowledge and collaborate with the tools and data your organization already employs to accomplish the task.

The conversation may evolve into a document, an application, or a workflow that continues executing the task. Agent workspaces are intended for use by all members of your organization. You engage with them through your browser, eliminating the need for technical expertise or terminal proficiency. A workspace integrates agent sessions, persistent state, outputs, files, resource access, and an isolated runtime where the agent can execute code.

It is preloaded with the curated context and skills accumulated by your team or organization. No longer does each task require reinventing the wheel—should someone on your team have optimized a process, everyone stands to gain. Individuals no longer must repeatedly explain the same procedures, terminology, and best practices to a model for every new task.

Request a workspace to investigate a subject using company context and the resources you make accessible. The agent can generate code to search, filter, join, and analyze information instead of importing an entire dataset into the model's context window. A workspace can transform its research into a document, presentation, or spreadsheet that you can subsequently edit.

These outputs need not be static files; they can remain linked to live data, be updated as sources change, and can be exported to familiar formats or services like Google Drive. When a document or spreadsheet falls short, the agent can construct an application with its own interface, logic, and state. The application can leverage connected company resources and facilitate collaborative work among multiple users.

Not every undertaking requires a comprehensive agent session. Many tasks comprise a predictable sequence of steps with a limited scope for judgment. A workspace can transmute these tasks into largely deterministic workflows, executing predictable steps via code while employing a model only where its contribution is warranted. Workflows can be initiated on demand, scheduled, or in response to an event within a connected system.

Cloudflare OS grants agents and applications restricted access to systems of record through Gatekeepers (a topic elaborated upon in the security section below). It also enables integration with MCP servers already utilized by your organization via MCP Server Portals. As individuals begin experimenting with AI in professional settings, a common initial demand is for API keys to access company systems.

This is understandable: AI lacks utility in a work environment without access to the systems employees employ. However, distributing API keys to individuals and agents is perilous and does not scale. Keys often convey extensive, indefinite access, posing challenges in constraining, sharing securely, and auditing. MCP offers a superior mechanism for agents to utilize these resources.