I run a small software studio that builds client work heavily using Claude. The problem I kept running into: clients would ask what the AI actually cost, and I had no clean answer. Token-count estimates drifted from the real bill, and nothing told me which project was spending what. I looked at SDK wrappers first. The problem is they only catch calls you make directly — not ChatGPT, not Gemini Code Assist, not anything going through a tool you don't control. A local proxy captures everything on the wire without touching a single line of application code. So Halton Meter intercepts outbound LLM API traffic, attributes each request to a project (env var, working directory, or process tree), and writes the cost to a local SQLite database. Nothing about how you call the API changes. `pipx install halton-meter` then `halton-meter init --apps` is the entire install. Supports Claude, OpenAI, Gemini, and Grok — including OAuth surfaces. No cloud. No tracking. Your data stays on your machine. Happy to answer questions on how the attribution chain works or how it handles HTTPS interception.

Halton Meter

Per-project LLM cost tracking. Local. No code changes.

Per-project LLM cost tracking. Local. No code changes.