Quickstart
Three ways in. Pick the one that matches what you have today; a project can use all three, and they all land in the same experiments, logs and release gate.
| You have | Start with | Time | You get |
|---|---|---|---|
| Code that calls your agent or model, and a few example inputs | SDK: an Eval in TypeScript or Python (in development) | About 5 minutes | An experiment with scores per row, the change against your baseline and the release gate's verdict |
| An app in production (or staging) that calls OpenAI, Anthropic or anything with OpenTelemetry | Logs: OpenTelemetry, or the AI Gateway (in development) | About 2 minutes | Every model and tool call in Logs, with cost, latency and errors; online scores once you add a rule |
| An agent that calls tools and changes data, in a GitHub repository | Repository sandbox | About 30 minutes | Your whole agent run in a fresh virtual machine under normal, fault and attack conditions, with the data checked afterwards and a verdict backed by evidence |
Before any of them you need a Preworld Review account: open the invitation link you were sent, or sign up with an activation code (email admin@nice2.io for one). Then create a project ("New project" in the sidebar): one project per agent or app.
Not sure?
Start with Logs if the app already runs somewhere, with the SDK if you're still building it, and with the repository sandbox when what matters is what the agent does (the tools it calls, the data it changes, how it behaves when a tool fails or a document tries to hijack it). Targets compares every way to connect.
What happens next
- Turn traces and failures into dataset rows, pick or write scorers, and compare experiments.
- Set the release gate so every version gets "can ship" or "failed the release gate", in CI too.
- Add online scoring and alerts for production.