How it works

Six steps, one request, no code beyond a base-URL swap.

1. 1

   Swap your base URL

   Point your existing OpenAI or Anthropic client at our endpoint. Nothing else about your code changes - same request shape, same response shape, plus a receipt.

2. 2

   Triage decides what the request needs

   One fast, typed decision call - Jev when you've configured it, our own heuristic engine otherwise - answers three things: what kind of task is this, which model tier can handle it, and how much verification does the answer deserve. That last answer is what keeps the expensive checks rare.

3. 3

   The cheapest adequate model answers

   Routed through Vercel AI Gateway or OpenRouter at provider list price - whichever is configured, zero markup either way.

4. 4

   The verification ladder runs, cheapest first

   Schema and rule checks always run, for free. Groundedness runs when you supplied context. Cross-model and self-consistency only run when triage decided the request was critical and uncertain enough to be worth the extra cost.

5. 5

   A failed check escalates once

   The strong model your project declares answers instead, gets re-checked, and the receipt records exactly why the first answer wasn't good enough.

6. 6

   You get the answer and the receipt

   Every response carries what it cost, what your declared baseline model would have cost for the same tokens, which checks ran, and what each one found.