Use OpenLLM when an operator needs to test an agent workflow against a self-hosted or controlled open-model endpoint before relying on a hosted model provider. The workflow is to install OpenLLM, choose a supported model and required hardware, start an OpenAI-compatible server, point the agent runtime or SDK at the local base URL, run a fixed set of prompts or task cases, then compare quality, latency, cost, and failure modes against the baseline provider. Invoke this instead of simply running a model server when the agent team needs a bounded compatibility and evaluation pass before switching model backends. Record the model name, endpoint URL, hardware assumptions, prompt set, and observed regressions so the decision is reproducible. The scope boundary is serving and evaluating one model endpoint for a defined agent workflow. It is not a generic model-serving platform card, a BentoCloud deployment guide, or a blanket recommendation to replace production inference.