Claudoscope
macOS menu bar app for Claude Code session analytics
Hey, I'm Liran. I built Claudoscope because I found my database password in a Claude Code session file and realized nobody was watching what these agents store on disk.
It's a free macOS menu bar app that reads your local Claude Code session files and gives you a dashboard. Cost per session, token analytics, secrets scanning, config linting. Native Swift, fully offline, no telemetry, open source.
Why
I use Claude Code daily. The agent reads files from your projects as part of normal work, and everything it reads ends up in JSONL logs under ~/.claude/projects/. If it opens a .env file, those credentials now live in plaintext on your machine. Nobody checks.
On top of that, there's no way to answer "which sessions are expensive?" or "is my prompt cache actually working?" The API dashboard doesn't break anything down per session. I wanted that visibility, so I built it.
Where I am
I shipped the first version about a month ago. Launching on Product Hunt tomorrow. 590 downloads so far, all organic from GitHub and Homebrew. Distributed via Homebrew cask. Solo dev, nights and weekends.
No revenue, no paid tier, not planning one. It's fully open source and I want to keep it that way. 590 people have it installed and I want to hear what's broken or missing.
Stuff I learned from my own data
Building this taught me more about my Claude Code usage than months of actually using it.
Short sessions are expensive. Each one pays full input cost with no prompt cache benefit. I thought my big multi-hour refactors were the costly ones. They weren't. It was the fifty quick questions throughout the day, each loading full context and throwing it away.
The config linter was an accident. I wrote a quick check for my own CLAUDE.md, then ran it on a colleague's and found his was 4,000 tokens. That's 10% of context window eaten by instructions before the agent even starts. Now the linter has 19 rules and one of them caught a silent parser-breaking bug in a skill config that nobody had noticed.
Prompt cache hit rates vary more than I expected. Some of my projects were sitting at 60-70%. That means 30-40% of input tokens billed at full price every message.
What I want
Feedback. If you use Claude Code on a Mac, try it and tell me what sucks. If the secrets scanner finds something in your logs, well, now you know.
I'll post updates here as things develop.
brew tap cordwainersmith/tap && brew install --cask claudoscope
website: https://claudoscope.com/
Comment