claude-token-audit

scripts/token-audit.mjs is zero-dependency Node 22+ and reads only local files. It takes about 4 s for 200 MB of logs.

node token-audit.mjs                    # last 30 days, all projects → token-audit.md
node token-audit.mjs --days 7
node token-audit.mjs --project myapp    # project folders containing "myapp"
node token-audit.mjs --out ~/usage.md

It prints the top recommendations and writes a full markdown report:

• totals (fresh input, cache writes, cache reads, output, cache hit rate) and the API-equivalent cost from Claude Code's own per-session counter
• By project, Heaviest sessions (with peak context size), By model, By day
• Tool output fed back into context: which tools flood the context window (Read, Bash, MCP…)
• Biggest tool outputs: the single results over ~5k tokens, with a preview so you can see what it was
• Loaded into every session: CLAUDE.md sizes, MCP server count, description size of the skills in ~/.claude/skills and in installed, enabled plugins

Reading the results with the user

• One API response is logged across several lines with the same usage. The script de-duplicates by message ID, so totals are not inflated.
• Subagent (Agent tool) transcripts in <session>/subagents/ are included and counted under the session that launched them. Peak context is the main thread's.
• Cache reads cost ~10% of fresh input. A high hit rate (90%+) is good; long sessions are still expensive because every turn re-reads the whole context.
• Recommendations, in order of usual impact: one task per session (/clear), /compact before continuing long work, trim noisy command output (| tail -50, -q), read files with offset/limit, use Sonnet/Haiku for routine work and subagents, remove unused MCP servers and skills, keep CLAUDE.md short.
• "Cost" is the API-equivalent number Claude Code computes. On a subscription the user isn't billed that, but it tracks how fast they burn through limits.
• Token counts for tool output are estimates (~4 characters per token). Request totals come straight from the API usage fields.