claude-dash
A zero-dependency Python tool that reads Claude Code's session transcripts and reports which model actually served each request, including subagent sidechains.
Python MIT licensed
When you run an AI coding agent all day, the model badge in the UI is a claim, not a receipt. claude-dash reads the JSONL transcripts Claude Code writes to disk and reports which model actually handled every session, including the subagent sidechains that are easy to miss.
It is plain Python 3.8+, no dependencies, MIT licensed. The parser streams each file in a single pass with a substring pre-filter before any JSON decoding, which keeps it quick: 63 sessions spanning hundreds of megabytes of transcript in about 0.7 seconds. Alongside the CLI there is a small local web dashboard for browsing the same data.
It exists because I wanted the answer to a simple question and the honest way to get it was to measure. That instinct, verify what the tooling tells you, runs through most of my recent work.