I built TokenTab because the OpenAI dashboard tells you what you spent but not which of your customers caused it. It is a proxy: you change your SDK base URL to /proxy/openai/v1 (Anthropic and Gemini paths too), keep your provider key in an X-Upstream-Key header, and add X-Customer and optionally X-Feature per request. The response passes through unchanged; only the usage block is read and priced against a bundled rate table. GET /report?group=customer|feature|model returns rows sorted by cost, so the first row is the answer. POST /meters pushes yesterday's tokens per customer into Stripe Billing Meters at 02:00 UTC with idempotent identifiers. Limitations in v1: non-streaming requests only, unknown models record tokens at zero cost, bodies are capped at 4MB, and request bodies are not stored. The first 10,000 requests a month are free, then $29 for a million and $79 for ten million. I would like to know whether grouping by feature matches how you think about margin.

TokenTab

Per-customer LLM cost attribution by changing a base URL

Per-customer LLM cost attribution by changing a base URL