Hi Vercel community! I’m building BabyLovable, an open-source, multi-user coding agent that turns a description into a Next.js app in an isolated workspace, with a live preview and iterative edits through chat.

GitHub: https://github.com/1nFrastr/baby-lovable Demo: https://baby-lovable.vercel.app/

The architecture combines several Vercel tools:

• AI SDK v7 and WorkflowAgent orchestrate the builder agent and its tools. Workflow DevKit gives long-running work durable steps and recovery. • Vercel Sandbox provides isolated workspaces for generated projects and their development previews; Daytona is also supported. • AI Gateway routes model calls, while Next.js hosts the chat and application APIs. • Vercel Connect is used for delegated third-party service connections. We are also working on publishing generated apps to users’ Vercel Teams through Connect; the Vercel OAuth API permissions and live publish flow are still awaiting access and verification.

One interesting challenge has been coordinating sandbox and preview lifecycle across concurrent serverless requests. The app reconciles desired workspace state with leases and version checks, while Supabase stores session state and Freestyle Git preserves source across ephemeral sandboxes.

I’d love feedback from people building durable coding agents: how do you handle workspace recovery, preview readiness, and verification after an agent edits files? Issues and contributions are welcome!