The Cloudflare AI team has open-sourced VibeSDK, a full-stack platform that enables developers to build and deploy AI applications with a single click on Cloudflare’s network. The platform is designed to let teams create their own AI-powered application builders without needing to integrate extensive infrastructure from scratch.
VibeSDK is a production-ready reference implementation provided under an MIT license. It includes a complete technology stack with a React and Vite front end, a back end built on Cloudflare Workers and Durable Objects, and data management using D1 and Drizzle.
The application creation workflow with VibeSDK
The process of building an application with VibeSDK follows a structured, AI-driven workflow.
- A user provides a description of the application they want to build.
- An AI agent generates the necessary code and files within a secure, isolated sandbox environment.
- The sandbox automatically installs project dependencies and starts a development server, exposing a public preview URL for the user to see the live application.
- Logs and error messages are streamed back to the generative agent, creating a feedback loop that allows the AI to perform iterative debugging and apply fixes.
- Once ready, a separate deployment sandbox publishes the application, ensuring it is provisioned as its own tenant-isolated Worker with a unique URL.
AI model flexibility and management
By default, the platform uses Google’s Gemini 2.5 models for planning, code generation, and debugging. All calls to large language models are routed through the Cloudflare AI Gateway, which provides several advantages. This architecture allows for seamless switching between different model providers like OpenAI, Anthropic, and Google. It also enables response caching, detailed observability into token usage and latency, and comprehensive cost tracking.
Security and multi-tenant architecture
The system is designed to securely handle untrusted, AI-generated code. Every build process runs within an isolated container or sandbox with strictly controlled network access. The production deployment architecture is inherently multi-tenant, achieved through per-application Worker isolation and resource usage limits. This model is designed to scale to support thousands or millions of individual user applications without the risk of cross-tenant data interference.
Deployment and project exporting
Cloudflare provides a live demonstration of VibeSDK, which includes a one-click button for users to deploy their own instance. Once an application is generated, users can export the project to a personal Cloudflare account or a GitHub repository. This functionality allows developers to continue working on their projects independently or integrate them with their own CI/CD pipelines.
A new pattern for AI-powered development
VibeSDK formalizes a development pattern known as “vibe coding,” which shifts the focus from manual coding to supervising generative AI agents. It provides a reproducible architecture that features safe code execution, integrated feedback loops for iterative development, and access to global deployment.
For organizations looking to create AI-powered builders, the platform offers a baseline that can be forked and specialized, potentially saving weeks or months of integration work.
