Your AI bill doubled.
Nobody can say why.
Your coding agents already write a log of every call they make. llmreceipts reads those logs and hands you an itemized receipt — which dollars were wasted, what caused it, and the exact line to fix.
npx llmreceipts scan
Read-only · runs locally · never sees your API keys
One command. One honest number.
A real receipt from a 14-engineer team's first scan. Every line is computed from the token counts already in their logs — not estimated, not sampled.
View as table
| Cause | Wasted | Share |
|---|---|---|
| Cache misses on a stable prefix | $612 | 37.4% |
| Duplicate & repeated calls | $364 | 22.2% |
| Over-powered model for the task | $298 | 18.2% |
| Context bloat (never read) | $187 | 11.4% |
| Retry storms | $121 | 7.4% |
| Abandoned turns | $55 | 3.4% |
| Total | $1,637 | 100% |
The gap is the product.
The top line is what you paid. The bottom line is what actually moved work forward. Everything between them is the money llmreceipts hands back to you.
View as table
| Month (cumulative) | Billed | Productive | Wasted | Waste % |
|---|---|---|---|---|
| Feb 2026 | $1,180 | $732 | $448 | 38.0% |
| Mar 2026 | $3,120 | $1,936 | $1,184 | 37.9% |
| Apr 2026 | $5,730 | $3,544 | $2,186 | 38.2% |
| May 2026 | $8,970 | $5,528 | $3,442 | 38.4% |
| Jun 2026 | $12,730 | $7,829 | $4,901 | 38.5% |
| Jul 2026 | $16,948 | $10,410 | $6,538 | 38.6% |
No proxy. Nothing in your critical path.
Every other cost tool asks you to route production traffic through their servers. llmreceipts never touches a live request — it reads files that are already on disk.
Point it at your logs
Claude Code, Cursor and Codex each write a session transcript locally, with exact input, output, cache-read and cache-write token counts per call. That's the raw material.
We grade every call
Deterministic rules — prompt hashing, prefix-stability analysis, task-to-model matching. No guessing, no LLM judging your LLM. Every dollar traces to a call ID.
You get a ranked fix list
Sorted by dollars recoverable, each with a file and line. Ship the top two and the next receipt proves it worked.
Six ways teams burn money without noticing.
All six are computable from the logs. That's the bar — if we can't prove it, we don't bill you for finding it.
Broken prompt caching
A timestamp or shuffled tool list at the top of your prompt silently invalidates the cache on every single call. Usually the single biggest line.
The same call, twice
Prompt-hash collisions inside a short window — agents re-asking a question they already paid for.
Wrong model, right answer
Frontier-model pricing for commit messages, lint fixes and file renames. Same result, a tenth of the cost.
Context nobody read
Whole files pulled into the window, billed in full, never referenced in the response. Also: the same file re-read nine times in one session.
Retry storms
Error handling that quietly bills you five times for one failure — and the agent loops that turn 6 calls into 40.
Abandoned output
Tokens generated, charged, then thrown away when a turn was cancelled or a tool result was truncated unread.
Your provider's dashboard tells you what you spent.
It will never tell you which line of your code is wasting it. That's the whole product.
We never hold your keys
No base-URL swap, no gateway, no secret handed to a vendor you met last week. Read-only, on your machine.
Zero latency added
We're not in the request path, so we can't slow you down or take you down. Uninstalling changes nothing about how your app runs.
Fixes, not dashboards
Not another chart of your spend going up. A ranked list of things to change, each with the dollars it gives back.
Find out what your number is.
We're onboarding a first group of teams spending $2k+/month on coding agents. First scan is free, and it takes about a minute.
No card, no call. We'll send the CLI and a sample report.