mcpField
A
auditing-agent-token-usage

auditing-agent-token-usage

github.com/bostonaholicAgent workflows266 installsUpdated today
Repository

About auditing-agent-token-usage

auditing-agent-token-usage runs a deterministic script to analyze agent token-usage data, then guides interpretation of its results. It reads Claude Code transcripts and Codex rollouts and distinguishes Claude Code proxy measures from recorded Codex token counts.

Key features

  1. 01Run the token-usage audit script
  2. 02Set an analysis window
  3. 03Report available data sources
  4. 04Identify evidenced usage costs
  5. 05Recommend supported changes

Use cases

  1. 01Reviewing agent usage over a chosen period
  2. 02Comparing available Claude Code and Codex usage data
  3. 03Finding usage patterns supported by audit output

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/bostonaholic/skills --skill auditing-agent-token-usage

After installing, type/auditing-agent-token-usagein Claude Code to confirm it loaded.

Bundled filesagents/openai.yamlscripts/audit-token-usage.sh

Details

From the project README

Auditing agent token usage

Run the deterministic audit, then explain its results. The script is the measurement; do not substitute your own log reading for it.

"<skill-dir>/scripts/audit-token-usage.sh" <days>

<days> is a positive integer, default seven unless the user names another window. The script needs jq and reads Claude Code transcripts and Codex rollouts under the home directory. On a nonzero exit, relay stderr and stop.

Analyze

  • State the window and which sources had data. Claude Code sections, [1] to [5], measure transcript proxies such as turn counts and character sizes. The Codex section, [6], uses recorded token counts.
  • Identify the largest evidenced costs: expensive model usage, inherited subagent routing, long sessions, compactions, context re-reading, oversized tool results, and configured MCP servers absent from section [5].
  • Recommend only changes the output supports: at most three, ranked by expected impact, each citing the count or percentage behind it.
  • Never present Claude Code proxy values as measured tokens.
  • When a source is absent or the script fails, report the missing data or prerequisite. Do not infer its usage.

Comments

Markdown supported