Sessions flow through Corveil.
Claude Code, Codex and Cursor connect to the Corveil gateway today. Every request is logged with its model, tokens and cost, under your team’s keys, budgets and guardrails.
Tool-Call X-ray
Drop in your Claude Code or Codex transcripts. Your browser finds the failed, repeated and looping tool calls and the context your agents paid to resend, prices it from the token usage in the files, and writes the patch that stops it. It also flags the API keys, secret files and personal data your agents sent to the model.
Drop transcripts or a folder here
Claude Code .jsonl sessions or Codex rollouts. A week of sessions is a good sample; hundreds of files are fine.
What happens to your data
Your transcripts were already on your disk before the scan; the X-ray only reads them. If it flags a secret, rotate it: it was sent to the model provider. See our privacy policy.
Where your transcripts are
Every coding agent session is saved as a JSONL file. Point the picker at the folder, or open it in your file manager and drag files in.
On a Mac, the folders are hidden: in the file picker press ⌘ ⇧ G and paste the path, or ⌘ ⇧ . to show hidden folders.
~/.claude/projects/~/.codex/sessions/find ~/.claude/projects ~/.codex/sessions -name '*.jsonl' -mtime -7The last one copies a week of sessions into ~/Desktop/agent-week so you can drop that one folder in. Dropping the whole ~/.claude/projects folder keeps the per-project breakdown.
Your X-ray
Every finding with its fix, the secrets and personal data (masked), the team projection and the patch, in one Markdown file built in this browser. Leave your work email and it downloads right away.
Your report is downloading as xray-report.md.
Thanks. We’ll get in touch to set up the daily team X-ray for your team.
What happens to your data
Your transcripts stay in this tab until you clear the results or close it. See our privacy policy.
Found in this browser, like the rest. A secret listed here was sent to the model provider and still sits in the transcript on your machine, so rotate it. Values are masked: at most their first 4 characters and their length (secret files are shown by path). None of this is priced or counted in the spend above.
| Project | Spend | Avoidable |
|---|
Generated only from what these sessions show. Review it, then commit it where your agents read instructions.
For the people driving the agents
Ifengineers work like this, about a month is avoidable.
Fixed rules, not a model: the same files always give the same result. Spend is every model request in the files at list API prices for the model named in the transcript.
Wasted tool calls. A call counts once, even when several rules match it. Its cost is the model request that made it: that request’s input, cache and output tokens, split evenly across the tool calls it made. Oversized results are also charged the cache reads they cause on later requests (up to 100, or until the context is compacted). A shell command that merely exits non-zero, like a search with no match or a failing test, is normal work and is not counted.
Context rebuilt after idle. A request whose previous request was longer ago than the cache lifetime it wrote with (60 minutes for 1-hour cache writes, 5 minutes for 5-minute ones), and that wrote at least 20k tokens to cache: the cache had expired and the whole context was written again. Cost: the cache-write price of those tokens minus what reading them from cache would have cost. This is an upper bound: once a cache has expired some reload is unavoidable, but /clear or a fresh session avoids most of it.
Context past 200k. A request carrying more than 200k tokens of context resends all of it every turn. Cost: the tokens past 200k at the cache-read price, the cheapest way those tokens can be billed. That is what a session split at task boundaries would not have resent.
The part of a request already charged to a wasted tool call is left out of both context costs, so no dollar is counted twice.
Secrets and personal data. Counted, not priced, and kept out of every figure above. The checks look at the commands the agent ran, what tools returned, what it wrote into files and the prompts, up to the first 32 KB of each. Secrets are known key formats (Anthropic, OpenAI, Corveil, GitHub, GitLab, Slack and AWS keys, JWTs, Bearer and Basic credentials, private keys), passwords in URLs, and values assigned to names like API_KEY, *_TOKEN or --password that have letters and digits. Secret files are .env files (not templates such as .env.example), ~/.ssh, ~/.aws/credentials, .pem and .key files, .npmrc and .netrc, read with the read tool or printed by a shell command. Personal data is email addresses (not no-reply or example addresses), Social Security numbers written 123-45-6789 that the SSA could have issued, and card numbers with a card network’s prefix that pass the Luhn check. Each value counts once per command, result, file write or prompt it appears in. Values are never shown in full, copied or sent.
What this page sends: to Google, usage events saying that the page was used and which of its buttons were clicked, such as that a scan ran, the patch was copied, the report was downloaded or the results were cleared, alongside the standard page view every page records. None of these events carries anything about your files or results: no counts, totals, coding agent, categories or findings, and never file contents, file or project names, commands, or model output. Only if you sign up, for the report or the daily team X-ray, does it send Corveil your email and which of the two you asked for. If you ran a scan, you can choose to include your scan summary (rounded totals: which coding agent, how many files, the waste share and avoidable-dollars bands, the biggest category, your team size band and a band for how many secrets, secret files or pieces of personal data were found); it is off unless you tick the box. The report itself never leaves your browser. See our privacy policy.
For your whole team
This page looks at one laptop, once. With Corveil, your team’s coding agents send their model traffic through one gateway, so the whole team’s usage is in one place. Corveil runs the team X-ray on top of it: the same rules, run every day, with each fix filed as a ticket. We set it up with you.
Claude Code, Codex and Cursor connect to the Corveil gateway today. Every request is logged with its model, tokens and cost, under your team’s keys, budgets and guardrails.
A Corveil Worker runs these rules over your team’s sessions every day, ranks the waste in dollars, and traces each item to its cause: a stale CLAUDE.md, a missing flag, a sandbox rule, a session habit.
Each fix lands as a GitHub issue with the evidence and the patch attached, so the instruction, tool or permission is fixed once instead of paid for in every session.
Run the X-ray above to see what your team is paying for calls that go nowhere.
What a daily ticket looks like:
Leave your work email and we’ll get in touch to set it up for your team.
## Agent friction: Flag that does not exist (`gh pr view`) Found: 14 wasted calls, ~2.1M tokens (~$4.80) across 9 sessions Fix (Missing flag): agents pass `--body` to `gh pr view`, which does not accept it. Patch: CLAUDE.md - `gh pr view` has no `--body` flag. Check `gh pr view --help` before adding flags. Example ticket. The figures are illustrative, not from a real team.