How to Track Your Claude Code and Codex Token Usage

Short version: both Claude Code and Codex keep a local log of every session. Claude Code stores them at ~/.claude/projects/**/*.jsonl; Codex at ~/.codex/sessions/**/rollout-*.jsonl. Those files hold token counts, model names, and timestamps — never your source code. You can read them by hand, with a CLI, or with a visual dashboard.

Where Claude Code and Codex store your usage logs

Both tools write a local log file for every session, with no setup on your part. Here is where to look:

Claude Code writes one JSONL file per session, grouped into a subfolder per project (its name is derived from that project's path). Codex writes one rollout-*.jsonl per session inside dated subfolders.

What's inside the logs — and what isn't

Each line is one event. Every assistant reply carries a usage block recording the tokens that round-trip spent. You can read:

• ✦Token counts for four classes: input, output, cache-read, and cache-write
• ✦The model that answered (for example, a Claude or GPT model identifier)
• ✦Timestamps, a session id, the project folder name, and usually the git branch
• ✦An AI-generated session title and the message-type structure (used to tell real user turns from tool results)
• ✦But not your source-code files

The raw files are verbose — they also carry system prompts and message text. But to total your usage you only need the token and model fields, and a good tool touches only those. Token Forest, for instance, reads only the counts and metadata; it never copies your prompts, replies, or source code into its own data, and never uploads anything.

The four token classes, and why raw counts look huge

Cache-read tokens usually dominate the raw counts, yet they are the cheapest class per token. So looking at a raw token total alone, it's easy to be alarmed by an enormous number.

In Token Forest's own six-week sample, cache reads were about 57% of real cost while input and output together were only about 24%. In other words, a huge raw-token number doesn't mean a huge bill. Judge spend by estimated cost, not by raw token totals.

Method 1: read the logs by hand

1. 1Open Claude Code's log folder at ~/.claude/projects/. You'll see one subfolder per project; inside are the .jsonl session files.
2. 2Open Codex's folder at ~/.codex/sessions/. Sessions are saved as rollout-*.jsonl inside dated subfolders.
3. 3Open any file in a text editor. Every line is one JSON event; assistant replies include a usage object with the four token counts.
4. 4To total them without reading every line, pipe a file through a tool like jq. Handy for a spot check — but tedious across hundreds of sessions.

# Show the token usage from one Claude Code session
jq 'select(.message.usage) | .message.usage' \
  ~/.claude/projects/<project>/<session>.jsonl

Method 2: use a command-line tool

If you'd rather not parse JSON yourself, a CLI can total it up. ccusage (open-source, by ryoppippi, on npm) reads the same ~/.claude logs — and now Codex logs too — and prints daily, monthly, and per-session cost tables in your terminal. Run it with npx ccusage. It's a well-loved, independent tool and a great fit if you live in the terminal.

Method 3 (the easy way): Token Forest's offline dashboard

If you want that same data as charts you can read at a glance — without touching the command line — install Token Forest. It reads the very same local logs, computes everything on your device, and shows three reconciling views. Nothing is uploaded; it works with the network fully switched off.

1. 1Download and install Token Forest (Windows 10/11 or macOS). It's a free public beta.
2. 2On first launch it scans your local Claude Code and Codex logs — no account, no API key.
3. 3Open the dashboard and read three views (below).

• ✦Growth — your usage drawn as a pixel tree that grows through stages, with days planted and 'this tree ≈ $X'.
• ✦Usage — daily / weekly / monthly charts, a per-model breakdown for both Claude Code and Codex, a 26-week heatmap, burn rate, and a per-project breakdown.
• ✦Chats — one line per conversation: title, project, prompt count, all four token classes, and an estimated cost.

The cost figures here are an offline estimate from a bundled price table, so they may differ from your actual Anthropic or OpenAI bill — but they're computed the same way every time, which makes them ideal for spotting trends.

Which method should you use?

Key takeaway: the logs are already on your machine — you need no API key and no internet connection to see your Claude Code and Codex usage. Read them by hand for a spot check, with a CLI like ccusage for quick totals, or with Token Forest for an at-a-glance dashboard. Whichever you pick, judge spend by estimated cost, not raw token counts.