Inspired By Frustrationmenu
Context Tracker · free macOS app
Is your repo Lean, Heavy, or Excessive?
↳ Claude Code loads instructions, memory and config before you type a word. Context Tracker measures exactly how many tokens of that startup context your repo costs.
Pick a local repository. The app runs one real headless Claude Code probe and reports the session total Claude itself measured — no guesswork, no estimate standing in for the verdict.
The three bands
< 30k
Lean
Startup context stays out of the way.
30–60k
Heavy
Worth a look at what loads before your first prompt.
> 60k
Excessive
Claude is spending real context on setup, every session.
The band is set by the startup token total Claude reports for the probe session — the same number every view on the dashboard hangs off.
how it works
One repo. One probe. One honest number.
↳ Context Tracker does not model your context window — it runs Claude Code and reads what actually loaded.
Claude's reported session total is authoritative. The category and file splits shown next to it are local estimates, and the app says so rather than dressing them up as measurements.
- Pick a local repository
Point Context Tracker at any folder you work in. Nothing is uploaded — the scan runs on your machine against your own checkout.
- One real Claude Code probe
The app runs a single headless Claude Code session in that repo. Not a simulation and not a token estimate of its own — a real startup, measured.
- Read the startup context total
Claude reports the session total, and that number is the verdict. The category and per-file splits beside it are local estimates of where that context came from.
What you get to look at
- Dashboard
The startup token total for this scan and the band it lands in — Lean, Heavy, or Excessive.
- Breakdown
Where the startup context appears to come from, by category. Local estimates, not Claude-reported numbers.
- Instruction Files
Which instruction files the repo carries and roughly what each one costs — the usual home of context bloat.
- Report
The scan written up in one place, for pasting into a PR, a ticket, or a context-engineering review.
- Trends
Plots the scans recorded on earlier runs, so a repo that keeps growing its startup context is visible before it hurts.
- Compare
A leaderboard across the repos you have scanned — which checkout costs Claude the most before you have typed anything.
Free, local, and quiet
- Free, no account
A notarized macOS DMG. Download, open, scan. There is nothing to sign up for.
- No telemetry
No usage reporting, no credential storage, and the contents of your instruction files are never stored.
- History stays local
Scan history lives in ~/Library/Application Support/ContextTracker/scan-history.json — the newest 200 scans, clearable from Trends → Clear Recorded Data.
Before you download
- macOS 14 or later
Apple-notarized build, so it opens without a Gatekeeper detour.
- Claude Code CLI, installed and signed in
The scan drives your own Claude Code install, so it has to be present and authenticated.
- Each scan uses your Claude allowance
One real headless session per scan. That is the price of a measured number instead of a guessed one.
Context engineering · Claude Code
Measure it before you trim it.
↳ Cutting instruction files is guesswork until you know what Claude actually loads at startup. Scan first, then decide what earns its place.
Free, notarized, and the download always serves the newest build.