Skip to content
ProductBlogOpen console
LLM Gateway

LLM Gateway

Every answer your agent gives costs money and takes time, and both are easy to discover only at the end of the month. The gateway is the single door all model traffic goes through, which is what makes those two numbers visible while you can still do something about them.

Calling a provider directly is a bit like giving everyone in the building their own front door key. It works right up until you want to know who came in.

Surface Answers
Providers Which vendors and models are available, and whose keys pay for them.
Budgets How much may be spent, over what period, by whom.
Guards Which models are permitted, and what a request may contain.
Usage How many calls, which models, how many tokens.
Spend What it actually cost.
Forecast What it is on track to cost.

You can run on the platform’s pooled capacity or supply your own provider credentials. Own keys mean your own rate limits and your own bill, which is usually what a larger workspace wants once traffic is predictable.

Credentials are stored encrypted and are never returned by the API after they are set. If you need to know whether a key still works, the gateway will tell you it is healthy. It will not tell you what it is, and neither will anybody else, which is the entire point.

They get conflated, so it is worth separating them plainly.

A budget is about money. It caps spend over a period and it does not care which model burned it.

A guard is about permission. It decides whether a particular request is allowed at all: which models may be used, and what may be sent to them.

A budget stops you spending too much on the right thing. A guard stops you doing the wrong thing cheaply. You want both.

Upstream rate limits do not always arrive wearing a name badge. A gateway can return a generic unavailable error when what actually happened is that your provider account hit its tokens-per-minute ceiling. If calls start failing in bursts and recovering on their own, suspect your provider’s throughput limit before you suspect the model.

Configuring providers and keys, setting budgets, writing guards, and reading usage against spend.