How to Use Vercel AI SDK with Cloudflare Durable Objects

Summary

This developer tutorial shows how to combine Vercel’s AI SDK with Cloudflare Durable Objects to build a more reliable interface for long-running AI requests. The author starts with a workout-template application whose original modal-based workflow could lose an in-flight or newly returned result when the user refreshed the page. A Durable Object is used as a long-lived, per-user execution context with dedicated SQLite storage and built-in WebSocket support. The application records sessions, prompts, and results in SQLite, then sends completion updates to the browser over open WebSocket connections. The article explains how Drizzle provides typed access to the Durable Object’s SQLite database and how `blockConcurrencyWhile` runs an idempotent schema setup before other requests use the object. A server route proxies WebSocket upgrades to the correct object, which is selected from the authenticated user ID. Prompt execution is started asynchronously, allowing the server to return a session ID immediately while the model request continues. On a session page, the application reloads stored history and subscribes to updates, so a refresh does not erase the workflow. The generated, schema-constrained workout data can then be displayed in existing form components and saved through the application’s normal endpoints. The author presents Durable Objects as a useful infrastructure layer for AI sessions, while describing Vercel’s AI SDK as a model-agnostic interface for making the underlying model requests.