In this livestream we walk through a working Forge LLMs app that rewrites a full blog draft in your own voice and streams the result into the UI while it is still being written: a concrete pattern for running long LLM jobs on Forge without hitting the invocation timeout wall.

We will go through its components: an async queue passing work to a long-running consumer, stream() plus Forge Realtime pushing tokens to the frontend and dynamic model resolution via list() instead of a hard-coded model name.

You'll learn how to:

• Wire up a queue, a consumer (also known as Async Events) and Forge Realtime so a long LLM job streams live into the UI instead of timing out.

• Resolve the model at runtime with list() and a preference order (defaulting to claude-opus-5), cache it and fail open, so a model deprecation never breaks your app

• Classify LLM failures into rate_limit, moderation, transient, validation and unknown, retry only what is retryable with bounded backoff, and keep the raw detail in Forge logs

• Record tokens, model as custom metrics and in Forge SQL, then display per-model usage details and running totals in the UI

Who it is for: developers building on Forge who want to use the LLMs API for real work, not just a demo. As always, everyone is welcome.

App sneak peek? Check it out here