Know when your Claude plan runs out. Before it does.

It sends counts and labels.

For each request: the token counts, the model name, the provider, the name of the folder the session ran in (never the full path), the time of the request, the provider's session id, a key that stops the same request, or the same upload, counting twice, and whether a sub-agent made the request. No cost is sent; the server works the cost out from the token counts when a record arrives. For each plan limit: the percent used, the window length, the reset time, the plan name and the time of the reading, with each window's kind and group, a severity word, the model's display name where a window is per model, and whether the window is active; for Codex, the limit's id and name, a line number, and a scrambled code for the reading's source, used only to keep readings in order. Each Claude limit reading carries a short scrambled code that says which Claude sign-in it belongs to. Each Codex limit reading carries the same kind of code for the Codex sign-in. Neither code can be turned back into your account.

It never sends your words, your code, or your files.

Not what you typed, not what the model answered, and not the contents of any file. To read your plan's limits it uses your Claude sign-in with Anthropic directly. That sign-in never reaches BoltGauge and is never written to a log. To read your Codex plan's limits it asks the Codex program on your machine, which talks to OpenAI with its own sign-in; the agent never reads that sign-in. To tell your Claude sign-ins apart it reads one value from ~/.claude.json, your account id, and scrambles it; for Codex it scrambles the account id Codex reports. Your email address and display name stay on your machine.