anthropometer  ·  amtr

The agentic debugger for Claude Code.

amtr attaches to a running agent session and lets you debug it the way you would a process: see the model's context window as a live memory map, step through what fills it and read the actual text, rewind to any turn, price every call, drill into each subagent's own window, and autopsy every compaction. Press R and it compiles a ground-truth PDF report of the whole session.

Works on Claude Code, Codex CLI and Gemini CLI sessions — every view, every key, one instrument (see Providers).

↑ replaying a session turn-by-turn: the context map fills, the composition shifts, the trend climbs.

▶ The autopsy: amtr was built in one 153-hour Claude Code conversation — so we pointed it at its own transcripts. 1,945 turns, 1.02 billion cache-read tokens, 3 compactions, $1,472 at API list price. The instrument dissects its own birth.

 ·   ·   ·   ·  Rust + ratatui TUI · Python engine · terminal-only

On a Mac? amtrino puts this in your menu bar — every live session a breathing dot, no terminal needed.  

---

AI usage

---

Why "debugger"

A monitor tells you how much. A debugger lets you look inside and move through time — and that is what these keys are: i INSPECT walks the context window segment by segment and reads back the bytes the model actually holds (⏎); ←/→ REPLAY re-renders every view at any past turn from the engine's own snapshot; ⏎ on an agent re-targets the whole instrument to that subagent's window (backspace returns); c opens a compaction's post-mortem — what was dropped, by category and by file. /context gives you one number. amtr gives you the whole picture, continuously: the context window as a live memory map, file access as a traffic seismograph, cache economics as a per-turn ledger, compactions as forensic events, and subagents as an economics table — every quantity labeled authoritative (read straight from the API usage records) or estimated, never blurred together.

It works on any session — Claude Code interactive or headless (claude -p), and Codex CLI — because every agent already writes a complete transcript. amtr just reads it in real time; new agents are a parser away.

Compared to token usage trackers

If you already run a token usage or cost tracker, amtr is not a replacement — it answers the question those tools stop at.