Build a custom AI operating system for your company
Ask for a meeting slide deck or a collaborative whiteboard, and the AI immediately writes and runs the code to create the app. The system generates a dedicated instance for each user request, allowing for safe sharing. Cloudflare has open-sourced the productivity tools it uses internally, enabling companies to build their own AI work environments.
Initial access stage after a full rewrite for version 2
The features are quite strong, but there is still much to refine, so it is classified as an early access stage. Cloudflare recommends that companies modify the repository to fit their specific situation and create their own 'company OS' rather than using it as is. Therefore, you must be fully aware of the current development status and limitations before adoption.
Method of running independent app instances for each user
Unlike existing SaaS services, this system creates a copy of the app and runs it in an independent sandbox whenever a user requests it. This is called a 'Gadget' and is completely isolated from other users' data, so information does not leak even if there are security vulnerabilities. If a user wants a feature that is missing, they can ask the AI agent to add it and directly modify the code. This structure moves away from the centralized software model, allowing users to freely modify apps to suit their needs.
External service access control managed by the Gatekeeper
When connecting to external services, a security layer called the 'Gatekeeper' intervenes to mediate access. The Gatekeeper handles OAuth authentication, restricts access to only the specific resources the user intends, and logs all actions. A particularly important point is that the approval process is handled asynchronously. When the agent performs a task, it can receive simulated results without waiting for approval and continue, while the user can approve or reject them in bulk later. This solves the problem of agents stopping due to approval waits, preventing interruptions in the workflow.
Preparation for local execution and external service integration
To test quickly locally, install pnpm, run the pnpm run-local command, and visit the http://localhost:8787 address. This method uses wrangler and workerd to run the full stack locally, but it is not suitable for production environments. You can also deploy to a Cloudflare account or to your own server using workerd. To integrate external services such as GitHub, Google, and Slack, you must set up OAuth client credentials according to the instructions included in each Gatekeeper package.
Technical prerequisites and contribution policy to check before adoption
This system is based on Cloudflare Workers and uses Durable Objects and Dynamic Workers as its core. Deploying to your own server requires an understanding of the workerd runtime, and related documentation and tools are currently under development. External contributions are limited, allowing only simple typo fixes or small PRs. This is because AI has made code writing easy, but code review and quality maintenance have become more important. If you are considering adoption, you need an understanding of the Workers ecosystem and patience for early-stage software.
