If you don't run your own LLM proxy, how do you know whats going on?

A developer argues that relying on AI providers' built-in usage dashboards is insufficient for real engineering decisions, since those panels lack time-series token data. The author's answer is a proxy you run yourself, with one endpoint across providers, logs of which key called which model and when, and virtual keys. He says that when he sees a cache miss during a coding session, he can pinpoint why it happened in 20 seconds by hand. The author dropped his Claude plan after finding that he can't use his purchased tokens in his harness of choice. A proxy server did not help either. Similar self-hosted options exist. LiteLLM, for example, is an OpenAI-compatible proxy you deploy and operate yourself. For a team building an AI product, a self-hosted proxy turns spend and model choice into data you own, not just a monthly statement. Per-key logs can show which service drives cost before the invoice arrives, and provider switching becomes a configuration change. The tradeoff is that you patch, host, and monitor the gateway yourself, so the infrastructure work becomes part of the product's operating load.