A self-hosted gateway with per-firm keys

Everything that would cost money is replaced by a mock that returns canned completions with a fixed usage block, so every number is predictable: 1,000 prompt and 500 completion tokens at the configured prices is $0.0105 a call. Everything else is real: the LiteLLM proxy, its Postgres-backed virtual keys, its budget checks and its router. The decision that matters most is pricing the mock models in the config: without input_cost_per_token and output_cost_per_token, LiteLLM records zero spend for an unknown model and the budget experiment shows nothing.