trackrecord

Your Claude Code track record. Lines added, languages, files, PRs shipped, sessions, streaks — parsed from the JSONL logs Claude Code already keeps on your machine.

npx trackrecord

Anthropic ships output analytics — lines accepted, PR tagging — only to Team and Enterprise customers via the Console dashboard. Cost trackers cover what you spent. Nobody serves individuals rigorous output metrics.

Teams get a dashboard for what Claude Code ships. trackrecord is that, for you — local, honest, free.

Zero network calls

The tool never touches the network, ever, in any code path. This is a hard invariant, not a privacy-policy promise:

• No fetch, http, net, dns, tls, or undici anywhere in the built output — enforced by an automated test that scans the published bundles (test/invariants.test.ts).
• No telemetry, permanently. It's on the project's never-list.
• Don't take our word for it: grep -rE "fetch\(|node:https?|undici" node_modules/trackrecord node_modules/@trackrecord after install, or read the source — it's small on purpose.

Your logs never leave your machine. Even trackrecord doctor (the debug report you might choose to paste into an issue) contains structure only — no code, prompts, or paths.

Commands

| Command | What it does | |---|---| | npx trackrecord | Terminal summary of your metrics | | npx trackrecord --json | Full schema object to stdout (schema/v1.json) | | npx trackrecord doctor | Anonymized structure survey of your logs, for bug reports |

All commands accept --dir <path> to override the default ~/.claude/projects/.

⚠ Your logs are evaporating (read this once)

Claude Code deletes transcripts after ~30 days by default. Your history — the corpus this tool reads — is on a rolling 30-day fuse unless you change one setting:

// ~/.claude/settings.json
{ "cleanupPeriodDays": 3650 }

This can't recover what's already gone, which is why every trackrecord command warns you when your corpus looks truncated. It's also why all stats are framed as "since <your first log>" — never "this year".

Methodology

The numbers are conservative by design. When the logs are ambiguous, trackrecord undercounts rather than guesses. (Anthropic's own Console analytics for teams applies the same posture — lockfile exclusion, conservative matching — which we take as validation of the approach.)

How lines are counted:

• A Write to a new file counts every line. A Write over a file we've seen before is diffed against the previous content — only genuinely new lines count. Rewriting a 500-line file to change 3 lines counts 3, not 500.
• An Edit is a line-diff of old_string vs new_string (Myers LCS). Added and removed are tracked separately.
• replace_all edits count once — the occurrence count isn't in the log, so we deliberately undercount.
• Lockfiles, *.min.*, and anything under dist/, build/, .next/, or node_modules/ is bucketed as generated and excluded from headline numbers.
• Lines are bucketed by extension: code, docs, config, styles, generated. The headline stat everywhere is lines of code added — docs and config don't inflate it.
• Blank lines count. Subagent output counts as output (subagent transcripts never count as sessions).

Scope statement: lines are counted from Claude Code's file tools (Write/Edit/MultiEdit/NotebookEdit). Code written via Bash heredocs, MCP tools, or other side channels is out of scope and not counted.

When trackrecord sees an unknown tool that looks like it writes files, it warns and asks you to run doctor and open an issue — it never guesses semantics into the count.

Sessions and activity: a session is a transcript file with at least one human prompt and one assistant reply — tool results, meta records, and compaction summaries are never counted as prompts; assistant turns are deduplicated by request id (streaming partials don't double-count). Token totals are labeled "API-equivalent value", never "spend" — they're what the usage would cost at API list prices, not what you paid.

Doctor

Hit a parsing bug, or numbers that look wrong? Run:

npx trackrecord doctor

It prints a markdown survey of your corpus structure — record types, field names, tool shapes, version range, anomaly counts. No message text, no code, no prompts, no paths (only .jsonl basenames). Paste it into a format report issue and the parser gets fixed for everyone.

FAQ

Why is my number smaller than tool X? Because we diff. Most counters credit every line of every Write — rewrite a file ten times and they count it ten times. trackrecord counts the delta, excludes generated files, and skips side channels it can't verify. Smaller and honest beats big and mushy; the whole point is a number you can defend.

Why don't my stats go back further? Claude Code deleted your older logs (see the retention section above). Fix cleanupPeriodDays today; the stats grow from here.

Is this a "Wrapped"? No — trackrecord reports delivery metrics over any range, any day of the year, not a year-in-review. (A shareable ship-card image lands at launch.)

How does this relate to ccusage / ccwrapped? Complementary: ccusage for cost, trackrecord for output. ccwrapped is orthogonal — personality and session vibes vs. delivery metrics. Run all three.

Why "trackrecord"? A track record is a proven history of delivered work — which is exactly what this shows. No cc- prefix soup, and not welded to one tool's name if the scope ever grows.

Schema

trackrecord --json emits a stable, versioned object — contract in schema/v1.json. Semver: additive fields bump minor, breaking changes bump major. Build dashboards on it.

License

MIT. Fonts: JetBrains Mono, in-repo for the upcoming card renderer under the SIL OFL (font bundling — not fetching — is part of the zero-network invariant; fonts are not part of the 0.1.0 npm package).