AI coding agent spend is scattered across every developer's machine: each agent keeps its own logs, in its own format, under its own plan. Tokenbase reads those logs where they already live and rolls them up into spend and usage by team, with Claude Code, Codex, Copilot, and Cursor normalized to one basis so the numbers actually compare. Everything is measured from the agents' own logs, not modeled from estimates. You can even use it to help you get hired!
No reviews yetBe the first to leave a review for Tokenbase
Maker
📌
Hey PH! Maker here.
AI coding agent spend lives on every developer's laptop. Claude Code, Codex, Copilot, and Cursor each keep their own logs, in their own format, under their own plan. So nobody can answer: what are we actually spending, and is it working?
Tokenbase reads those logs where they already live. One curl command installs a read-only collector, and everything rolls up into spend and usage by team, translated to one format so the numbers actually compare.
Two rules we won't break: every number is measured from the agents' own logs, never estimated. And your code and prompts never leave the machine.
Favorite detail: agents log their own rate limits, so you can see which sessions your plan cut off mid-work.
Ask me anything!
Report
finally something that reads the actual agent logs instead of guessing. saw claude code and cursor on the same chart without the usual apples to oranges mess, which is what i've been wanting for months.
Report
Maker
@aytennwxw yeah, that was the hard part. every agent counts tokens its own way, and a few do genuinely weird things, like counting reasoning twice or resetting their running total halfway through a session. so getting claude code and cursor onto one honest basis wasn't really parsing. it was going agent by agent and checking each one against the tool's own source until the numbers reconciled. that work is invisible when it's right. nice to hear it showed.
Report
honestly this looks super useful for keeping AI coding spend in check across the team. one thing that would be a nice add is a budget alert system, like setting monthly limits per team or project and getting a slack ping when you hit 80 percent. would save a lot of surprise overages
Report
Maker
@elelcizehr93509 love this one. it fits cleanly because everything already rolls up to the team, so an alert watches the team total and never a single person. not built yet. the part i keep chewing on is what you'd set the cap against. real dollars spent, or plan usage, meaning how close a claude or copilot plan is to maxing its window before it starts dropping sessions mid work. those are two pretty different alerts. which one would you actually reach for?
finally something that reads the actual agent logs instead of guessing. saw claude code and cursor on the same chart without the usual apples to oranges mess, which is what i've been wanting for months.
@aytennwxw yeah, that was the hard part. every agent counts tokens its own way, and a few do genuinely weird things, like counting reasoning twice or resetting their running total halfway through a session. so getting claude code and cursor onto one honest basis wasn't really parsing. it was going agent by agent and checking each one against the tool's own source until the numbers reconciled. that work is invisible when it's right. nice to hear it showed.
honestly this looks super useful for keeping AI coding spend in check across the team. one thing that would be a nice add is a budget alert system, like setting monthly limits per team or project and getting a slack ping when you hit 80 percent. would save a lot of surprise overages
@elelcizehr93509 love this one. it fits cleanly because everything already rolls up to the team, so an alert watches the team total and never a single person. not built yet. the part i keep chewing on is what you'd set the cap against. real dollars spent, or plan usage, meaning how close a claude or copilot plan is to maxing its window before it starts dropping sessions mid work. those are two pretty different alerts. which one would you actually reach for?