Hi everyone, I'm Evan. I use Claude Code and Codex every day, and a few weeks ago I actually added it up: 740 sessions and $4,054 of list-price usage in eight weeks, on one seat. I could not have told you what any of it was for. That is the gap. Vendor dashboards count tokens per user. They cannot say which repo, which squad, or which client a session belonged to, and on a subscription seat there is no dollar figure at all. So the two questions that actually matter go unanswered: which project is eating the budget, and who should be paying for it. Tallyhook installs with one command per machine, reads the session logs the agents already write, and turns each session into a ledger line: repo, developer, model, tokens, list cost. You group repos once, into teams if you report internally or clients if you bill, and every past and future session rolls up. One thing I did not expect while building it. Measured across 1,406 of my own session logs, about a third of spend was landing on the wrong project, because a session gets tagged with the directory you launched the agent in, not where the work actually went. Tallyhook resolves each touched file to its own repo instead, and splits a session across repos when it spans more than one. Metadata only. Prompts and code never leave the machine, and the collector is one dependency-free file you can read in ten minutes. Two ways to try it without an account: "npx tallyhook report" prints your cost per repo locally and uploads nothing, and there is a live demo at tallyhook.dev/demo with sample data for a five-developer agency. I would genuinely like to hear how other people handle AI costs in fixed-fee work. Happy to answer anything.

Tallyhook

What Claude Code and Codex cost, per team or client

What Claude Code and Codex cost, per team or client