Overview

agentmeter reads the session logs that coding agents already write to your disk and explains your usage: numbers from your own logs, each with the evidence behind it and a concrete next step to check. It targets Claude Code and Codex CLI.

Status: early development. Nothing is released yet; readers for Claude Code and Codex work, parts of this documentation (menu, configuration check, budget) describe planned behaviour.

What it reports

• A summary of requests, input, cache share, output and compactions over a period.
• Sessions and projects that account for most of the input.
• Findings: observations and hints from rules over the same data, such as large contexts, cache rebuilds and the size of the first request. Each states how sure it is.
• A data-quality section: observed period, files without usage, unreadable entries, unknown models.

Privacy

agentmeter only reads usage fields: token counts, timestamps, model, session and project ids and compaction events. It does not read prompts or file contents, does not touch the network and sends no telemetry.

Where it stops

• It cannot show the remaining quota of a subscription, because that is not stored locally. A later budget mode can estimate the API-equivalent value of your usage.
• Log formats of both agents are not official. Log formats lists what is read and where that is an assumption.
• Tokens are not comparable across vendors. Numbers are compared within one agent or in money.