Cloudflare OS lets you deploy a managed agent workspace that integrates your company's internal data and systems

Cloudflare opened a waitlist for a fully managed version of Cloudflare OS, an open-source workspace that lets AI agents access a company's CRM, support tickets, and internal docs, with setup requiring only a custom domain, Cloudflare Access policies, and a linked AI Gateway. The agents can now mount Git repos to search codebases, fix bugs, and push commits or open pull requests, and reach Gmail, Google Drive, and Sheets through a Google Workspace Gatekeeper Worker, with file export to .xlsx, CSV, PDF, Markdown, and HTML and .docx and .pptx support listed as coming soon. The source code's Gatekeeper, around line 617 in packages/workshop-shared/src/gatekeeper.ts, simulates the success of actions such as merging a pull request before human approval so the agent keeps working, and rejecting the first step discards every subsequent step built on that assumption.

Cloudflare OS lets you deploy a managed agent workspace that integrates your company's internal data and systems Cloudflare OS is essentially a collaborative environment where AI agents can access your CRM, support tickets, and internal docs to handle repetitive prep work—like synthesizing account data for customer meetings—without a human doing the manual digging. While the project is open source, Cloudflare is now opening a waitlist for a fully managed version that removes the need for internal teams to handle deployment and updates. How to set up a managed workspace? To get the managed version, you join a waitlist. Once granted access, the setup happens via the Cloudflare dashboard. You only need to provide three things: 1. A custom domain for the workspace. 2. Your specific Cloudflare Access policies. 3. The AI Gateway you want to link to the environment. If you prefer total control and don't want to wait for the managed service, you can deploy the project directly from their open-source repository into your own account, though you'll be responsible for the maintenance. What can the agents actually do now? The capabilities have expanded beyond basic document generation into actual technical and administrative workflows: - GitHub Integration: Agents can now mount Git repos. This means they can search through a codebase, fix bugs, add features, and actually push commits or open pull requests. - Google Workspace Access: Via the Google Workspace Gatekeeper a specialized Worker , agents can research Gmail threads, draft emails, and access specific folders or individual files in Google Drive and Sheets. - File Exporting: You can move work out of the workspace into .xlsx, CSV, PDF, Markdown, or HTML. Support for .docx and .pptx is listed as coming soon. The "Gatekeeper" logic and how it handles actions One of the most interesting technical details in the source code specifically around line 617 in packages/workshop-shared/src/gatekeeper.ts is how the system handles side effects like sending emails or merging code. Instead of letting an agent execute a command immediately, it goes through a Gatekeeper. To keep the agent moving through a multi-step task without stalling, the Gatekeeper "simulates" the success of an action before a human approves it. If an agent asks to merge a PR, the Gatekeeper tells the agent it happened—even if it hasn't. The agent then proceeds to queue the next steps based on that "fabricated reality." This means if you are reviewing a batch of agent actions and you reject the first step, every subsequent step the agent built on that assumption is automatically discarded. It prevents the agent from hitting a wall while waiting for human intervention, but it also means the agent's perceived state of the world is temporarily fictional until you hit commit. Next Muse AI agent's Data Security Risks Raise Red Flags → https://promptcube3.com/en/threads/9699/ All Replies （1） Want a live back-and-forth? Join the global AI chat room https://promptcube3.com/en/chat/ — login to talk. Connecting this to my CRM usually takes hours of API tinkering. A managed version is the only way to scale this.