Token Triage is Open Source. And It Just Audited Itself.
I built a free, local-only CLI for Claude Code logs. Auditing my own usage exposed a $0 pricing bug. Source and install command below.
My logs recorded $11,737 in API-equivalent cost in one month. 92.5% of the tokens processed were cache reads.
Install the free CLI (requires pipx and Git):
pipx install git+https://github.com/cursed180/Cursed_Token_Triage.git
token-triageThe first line installs it. The second runs it.
I've been running a small CLI called token-triage against my own Claude Code usage for a while. It reads the local event logs and tells you where your tokens actually went. It's public now, Apache-2.0, on GitHub.
The reason I trust it enough to release: before shipping it, I pointed it at a full month of my own fleet. 34,533 assistant turns, $11,737 in API-equivalent cost. The total isn't the interesting part. The shape is. 92.5% of every token processed was a cache read: the accumulated context, re-read again on every single turn. The prompts I typed by hand were 0.17% of the total. If you've ever wondered where agentic spend actually goes, it mostly isn't the part you're typing.
Then I caught a pricing bug while checking the tool's output. Fable 5, my most expensive model at 2x Opus rates, was metering at $0. Its model id doesn't carry an "opus/sonnet/haiku" family token, so it slipped past the pricing fallback silently. 8.7% of my calls, $1,213 that month, invisible. I only caught it because the total didn't smell right against what I knew I'd actually been running. Fixed it, added regression tests so it can't regress, shipped it.
Here's what I learned: a cost tool you don't sanity-check against reality will confidently hand you a wrong number. token-triage is stdlib-only, zero runtime dependencies, no network code anywhere in the package. CI fails the build if that ever changes. Local-only by design. If you can't ship raw session logs to a SaaS dashboard for compliance reasons, you still deserve to know what you're spending.
Full field study, with the by-model and by-project breakdowns: STUDY.md.
It's Apache-2.0: use it, fork it, tell me what it gets wrong.
FAQ
Is this my Max bill?
I report API-equivalent costs. On a flat Max plan, these describe usage at API prices. They are not your subscription bill or a promise of cash savings.
What will I see?
I use Token Triage to read local Claude Code event logs and show token use and API-equivalent costs by project, session, and model.
Do my logs leave my machine?
I built the CLI to run locally. It has no network code and no telemetry. Token Triage sends nothing off your machine. The pipx installation downloads code.
Why the paid links?
I offer a $1,500 burn audit for teams paying by token. You run the tool and send me its report. Agent Stack Setup is paid repo setup. Neither service is required to use Token Triage.