Claudon
Local-first analytics for Claude Code. Understand where time, cost and tokens go, surface bottlenecks and review model reasoning, entirely on your own machine.
Why Claudon?
Claude Code is fast until it isn't. Which turn burned ten minutes? Was it the model, a slow tool, or you being away? Did a failing tool call loop? What did the session really cost? The answers are already in ~/.claude. Claudon reads your session logs and turns them into one interactive HTML dashboard with a single command: uvx claudon --open. It is a small Python package that uses only the standard library. No telemetry, no network, no account.
What it shows
Every number is computed exactly from your own transcripts.
Where time went
Model vs. tools vs. permission waits vs. idle, per task, turn and subagent.
What it cost
Token usage multiplied by list prices per model generation, with your own rates via --pricing.
What went wrong
Ranked findings with a concrete fix each, from retry loops to expired prompt caches.
How hard it thought
Thinking time against thinking tokens, per model.
What to fix first
An optional AI summary and Ask AI chat, run by your browser's on-device model.
Safe to share
--redact strips prompts, paths, commands and project names before you attach a report.
Quick start
Run it with whichever tool you already have. Each one fetches the latest release on demand.
| Tool | Run once | Install globally |
|---|---|---|
| uv | uvx claudon | uv tool install claudon |
| pipx | pipx run claudon | pipx install claudon |
| pip | pip install claudon | |
| npm | npx @infenia/claudon | npm install -g @infenia/claudon |
| pnpm | pnpm dlx @infenia/claudon | pnpm add -g @infenia/claudon |
| Bun | bunx @infenia/claudon | bun add -g @infenia/claudon |
Bottlenecks it finds
Ranked worst first, each with the time and spend at stake and how to fix it.
Idle and permission waits
Time lost to being away, and to Read, Edit and Write calls waiting for your approval.
Failing tools and retry loops
Tools that keep erroring, and tasks stuck retrying the same failing step.
Repeated work
Identical reads and searches run again, and edit churn on the same file.
Context and cache
Context bloat past 400k tokens, prompt caches that expired after a pause, and a high cache-write share.
Interrupts and rejections
Interrupted turns, rejected tool calls, heavy thinking and subagent cost share.
Trends
Tasks with unusually many turns, cost concentration, and earlier vs. later tasks.
Runs in your browser too
No Python? Open the live app at claudon.infenia.com/app and drop in your .jsonl transcripts, or open your ~/.claude/projects folder. Everything is processed in the page. Inside Claude Code, run claudon --install-plugin once, then type /claudon in any session to generate and open your dashboard.
Private by design
Everything runs on your machine. The CLI makes no network requests, and neither does the report, a single self-contained HTML file. The optional AI summary and Ask AI chat use your browser's built-in on-device model, and stay hidden in browsers that don't support it.