What a live map of one Claude Code session showed me.
I typed two short prompts into Claude Code: give me a tour of this project, then run the tests and fix whatever fails. Together they came to about 41 tokens. At that moment, Claude Code was carrying 59,375 tokens of context.
That isn't a bug, and it isn't Claude Code being careless. It's how an agent works: every request carries the instructions, the tools, your project's CLAUDE.md, the skills it might use, and the whole conversation so far. But I'd never seen it laid out step by step while it happened, so I built a tool that does: Flightpath, a free, open-source Claude Code plugin. Here's what one seven-minute session looked like through it.
The session
A fresh Claude Code setup on a small sample project, a command-line tool for tide times. I gave the project a long CLAUDE.md and more skills than it needed, to see what Flightpath would make of them. Then a handful of ordinary requests: a tour, fix a failing test, read a build file, a subagent review, a model switch and a new flag, a build I refused, and a diff.
Seven minutes, 9 turns, 23 steps.
What it showed
Before Claude had answered anything, the context held about 55K tokens. Most of it is sent on every request: the system prompt, 19 tool definitions, my CLAUDE.md and the list of skills. Flightpath's Snapshot view breaks it down piece by piece, so you can see which parts are yours to trim.
One file read added 20,573 tokens in a single step. I asked Claude to read a source map in the build folder. It's a perfectly reasonable request, and it was the biggest jump in the whole session. The Tokens view lists the biggest rises and links each to the step that caused it, so a jump like that doesn't stay a mystery.
Switching model mid-session wrote 52.9K tokens of cache again. When I switched model with /model partway through, Claude Code couldn't reuse its cache, so the conversation was written again. Flightpath's coach flagged it at that step, with the reason and a link to Claude Code's documentation: pick the model at the start of a session when you can.
A subagent used 141.2K tokens to hand back about 1.6K. Subagents keep their work out of your main context, which is the point. But it's useful to see the trade: the main conversation got a short summary, and the subagent did the heavy reading on its own budget.
Flightpath itself added 0 tokens. It records through Claude Code's hooks and answers its own commands on your machine, so nothing it does reaches the model. Every session shows its overhead, so you can check.
What I took from it
Most of what Claude Code carries is the right stuff, and you wouldn't want it gone. The surprises were in the details: one file read that was far bigger than it looked, a model switch at the wrong moment, a start-up file that had grown. None of that shows in the terminal while it happens. Seeing it live makes the choices easy: how long a CLAUDE.md should be, when to switch models, and which folders Claude should read.
Try it
Flightpath runs alongside Claude Code and opens in your browser. Everything stays on your machine: it keeps file paths, never file contents. You need Node 24 or later. Then, inside Claude Code:
/plugin marketplace add dk-labs-dev/flightpath
/plugin install flightpath@flightpath
Your next session records automatically. Type /flightpath:open to watch it.
A 36-second look: https://youtube.com/shorts/IocHN_qt01E
The source is on GitHub under Apache 2.0: https://github.com/dk-labs-dev/flightpath
I'd like to hear what it shows on your sessions. Is your gap as wide as mine? And if Flightpath is useful to you, a star on GitHub helps other Claude Code users find it.


Top comments (1)
This is my area, since I wrote Tokenade to cut that part of the context. Your prompt is 41 tokens of the 59,375. The rest is the system prompt, the tool definitions and the memory files, and the model receives them anew at every turn. Measure the first turn and a late turn separately, because the gap between them is what you can move.