What is Capture?

Capture decides whether we store the content of your requests and responses, or only the numbers about them. It is off by default and it is your choice, per key.

Every request through your key is metered, so your usage log always shows the model called, the tokens in and out, the cost, the latency and the status code. What we do not keep is the conversation itself: the prompt you sent and the answer the model returned pass through and are gone. If a customer of yours sends you a legal document to summarise, no copy of that document sits on our servers.

We keep the request body and the response body alongside each log row, so when you open Logs in the dashboard you can click into a request and read exactly what was sent and exactly what came back. That is the whole feature: it turns the usage log from a bill into a debugger.

You also pick a retention window, 1, 7 or 30 days. When a body is older than the window it is deleted for good; the metered numbers stay.

While building. When a model answers something strange at 2am, the question is always "what did my code actually send?", and Capture answers it without adding print statements and reproducing the bug. Teams typically capture on their development key with a short window, and leave it off on production keys that carry end-user data.

Because whether to store people's words is a privacy decision, and it is yours, not ours. Different keys carry different traffic: your test key carries your own prompts, your production key may carry your customers'. Off by default means we never hold content you did not explicitly ask us to hold, which is also the answer you give your own customers when they ask what your AI vendor stores.