Codex Claude Code Token Dashboard

An H&K Brothers local-first usage analytics dashboard for Codex and Claude Code. It reads Codex JSONL rollout logs under ~/.codex/sessions/ and Claude Code project logs under ~/.claude/projects/, then turns them into prompt cost estimates, tool/file heatmaps, cache analytics, project comparisons, skill usage, session drill-downs, and rule-based efficiency tips.

Everything runs locally. No telemetry, no login, no remote calls for your session data.

Maintained as an H&K Brothers tool.

Screenshots

Codex overview

Claude Code overview

What this is useful for

• Seeing which prompts and sessions consume the most tokens.
• Comparing Codex and Claude Code usage across workspaces and projects.
• Spotting repeated shell commands, oversized tool results, and expensive turns.
• Understanding cached input versus fresh input from Codex token_count events.
• Estimating API-equivalent cost from editable model pricing.

Prerequisites

• Python 3.8 or newer.
• Codex and/or Claude Code installed and used at least once so local JSONL logs exist.
• A modern web browser.

No pip install, no Node.js, no build step.

Installation & Quickstart

git clone https://github.com/H-KBrothers/Codex-ClaudeCode-Token-Dashboard.git
cd Codex-ClaudeCode-Token-Dashboard
python3 cli.py dashboard

The command:

1. Scans ~/.codex/sessions/ and ~/.claude/projects/.
2. Builds/updates a local SQLite cache at ~/.codex/Codex-ClaudeCode-Token-Dashboard.db.
3. Starts a local server at http://127.0.0.1:8080.
4. Opens the dashboard in your browser.

Leave it running; it rescans every 30 seconds and pushes updates live. Stop with Ctrl+C.

Data Sources

Codex writes JSONL rollout files here:

OS	Path
macOS / Linux	~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl
Windows	C:\Users\<you>\.codex\sessions\YYYY\MM\DD\rollout-*.jsonl

The dashboard only reads these files. It never modifies Codex transcripts.

Claude Code writes JSONL project files here:

OS	Path
macOS / Linux	~/.claude/projects/<project-slug>/*.jsonl
Windows	C:\Users\<you>\.claude\projects\<project-slug>\*.jsonl

The top-bar switch lets you inspect Codex or Claude Code with separate visual themes and source filters.

To point at another location:

python3 cli.py dashboard --projects-dir /path/to/codex/sessions --claude-projects-dir /path/to/claude/projects --db /path/to/cache.db

Environment Variables

Var	Default	Purpose
PORT	8080	Local web server port
HOST	127.0.0.1	Bind address. Keep the default unless you intentionally want LAN access.
CODEX_SESSIONS_DIR	~/.codex/sessions	JSONL root to scan
CLAUDE_PROJECTS_DIR	~/.claude/projects	Claude Code JSONL root to scan
CODEX_DASHBOARD_DB	~/.codex/Codex-ClaudeCode-Token-Dashboard.db	SQLite cache path

TOKEN_DASHBOARD_DB and CODEX_PROJECTS_DIR are still accepted as compatibility aliases.

CLI Reference

python3 cli.py scan          # populate / refresh the local DB, then exit
python3 cli.py today         # today's totals in the terminal
python3 cli.py stats         # all-time totals in the terminal
python3 cli.py tips          # active suggestions in the terminal
python3 cli.py dashboard     # scan + serve the UI

python3 cli.py dashboard --no-open
python3 cli.py dashboard --no-scan

Change the port:

PORT=9000 python3 cli.py dashboard

Accuracy Notes

Codex token_count events include both cumulative totals and last_token_usage. This dashboard sums last_token_usage only; summing cumulative totals would over-count usage.

Codex reports cached_input_tokens inside the input total. The scanner splits that into cache_read_tokens and keeps only fresh input in input_tokens.

Claude Code transcripts use the original Claude Code message format. The scanner keeps those rows under source='claude' so they can be filtered separately from Codex.

Cost is API-equivalent. If you use Codex through a ChatGPT subscription, this does not mean you were separately billed that amount. Edit pricing.json when model prices change or when a private/internal Codex model slug needs a custom rate.

Privacy

The browser fetches JSON from 127.0.0.1, and all JS/CSS/chart code is served locally from this project. Use Cmd/Ctrl+B in the dashboard to blur prompt text before screenshots.

Tech Stack

Python stdlib for the CLI, scanner, SQLite cache, and HTTP/SSE server. Vanilla JS + ECharts for the UI.

Data flow:

cli.py -> token_dashboard/scanner.py -> SQLite -> token_dashboard/server.py -> web/

Development

python3 -m unittest discover tests

See CODEX.md for the architecture map and local conventions.

License

MIT. See LICENSE.