AgentCart AI

FOR DEVELOPERS AND TEAMS USING AI CODING TOOLS

AI credits disappeared? Find the run that used them

For unexpected credit or cost changes in Codex, Claude Code, OpenCode, OpenClaw, and Gemini.

Choose the app and its local records to find the unusual session, retry, background task, or model switch for free. No log-reading required.

Start free analysis See example
Free result the exact run why it happened what to do next
Free analysisSee the cause, amount, and evidence Optional · 1 USDCPay only for a shareable report
No sign-in Runs on this device Files are not uploaded Open-source code

START HERE · TWO STEPS

1. Where did you notice the unexpected usage?

Choose the app, then its local records. No account connection or API key is needed.
Choose the app
2. Choose the Codex usage folder We find compatible records automatically. You do not need to open or understand them.
RECOMMENDEDChoose records in this browser
Folder to choose ~/.codex/sessions Mac: press Cmd+Shift+G in the folder picker and paste this path. Windows: use %USERPROFILE%\.codex\sessions.

No usage history selected yet

or use automatic detection
COMFORTABLE WITH TERMINAL? Show the automatic audit command

Copy this into Terminal. It finds the local records and prints the answer.

npm exec --yes --package=github:zac343/agent-cost-auditor#v0.5.3 -- agent-cost-audit --auto codex
Runs only on this computer. Logs, prompts, and messages are not uploaded.
Advanced: session ID or pasted usage records
or paste records

Codex and Claude Code JSONL stream locally up to 2 GB. Codex uses last_token_usage deltas and marks exact ancestor replay. Claude Code assistant fragments are deduplicated by request, cache reads and writes are normalized into prompt totals, and main-session usage is separated from isSidechain background work. Other exports have a 10 MB combined limit. Exact duplicate rows require the same stable attempt or response ID and identical usage values; a root request ID alone never removes a fallback attempt. OpenCode sanitized session exports, OpenClaw transcripts, Gemini usage metadata including thoughtsTokenCount, and OpenRouter and Vercel AI Gateway generation JSON, CSV, JSONL, and NDJSON are recognized directly. Custom exports may use in/out, cache_read/cache_creation, observed_cost_usd, and cost_evaluable; an explicitly unevaluable cost is reported as unknown, never as $0.

YOUR RESULT

The cause, evidence, and next action appear here

Complete the two steps on the left, then start the free analysis.
10-SECOND EXAMPLE A failed retry used $0.0184

See the answer format before choosing your own records.

  • What happenedA failed request retained recorded spend
  • Proof3 records, $0.1054 total, 1 unpriced call
  • Next stepStop the retry and send the evidence to support
Example data only. It is never counted as a customer analysis.