Claude Token Dashboard

A local dashboard for what Claude actually costs — Code and Cowork together. Reads the JSONLs Claude Code writes to ~/.claude/projects/ and the audit logs Cowork writes to ~/Library/Application Support/Claude/local-agent-mode-sessions/, and turns both into per-prompt cost analytics, tool/file heatmaps, project comparisons, and cache analytics.

If you're on Claude Max ($100/mo) and use Cowork, the official tools won't tell you what your usage would have cost on pay-per-token rates, or whether the subscription pays for itself. This does — separately for each surface, or both stacked.

Everything runs locally. No telemetry, no login, no data leaves your machine.

Forked from nateherkai/token-dashboard (MIT) and extended with Cowork visibility, a source toggle, a plan-cost flip, and a redesign per Impeccable.

What's different from upstream

	Upstream	This fork
Sources	Claude Code only	Claude Code + Cowork
DB	messages, tool_calls	Same, plus a source column
Topbar	Brand + tabs	Brand + tabs + 3-way source toggle (all / code / cowork)
Cost card on subscription	API-equivalent figure (~$3,010), monthly fee in subline	Monthly fee leads ($100/mo), API-equivalent + value multiple in subline
Theme	Dark observability template	Light, OKLCH tinted-neutral palette
Color strategy	Multi-hue accent palette	Restrained — no chromatic accent, grayscale chart ramp
KPI display	Card grid	Hairline strip (hero-metric-template is an Impeccable absolute ban)

What this answers

• Am I getting my $100/mo of value out of Max? The cost card shows the multiple in plain English.
• Where are my tokens going? Per-prompt breakdown, file/tool heatmaps, project comparison.
• Why was that one session so expensive? Click any prompt → see the assistant response and tool-result sizes.
• Are my Cowork sessions burning cache reads? They probably are. The dashboard separates cache reads from billable input so you can tell.

Prerequisites

• Python 3.8+ (already on macOS).
• Claude Code, Cowork, or both — used at least once.
• A modern browser.

No pip install. No Node. No build step.

Quickstart

git clone https://github.com/VisionaireLabs/claude-token-dashboard.git
cd claude-token-dashboard
python3 cli.py dashboard

Scans both ~/.claude/projects/ and ~/Library/Application Support/Claude/local-agent-mode-sessions/, starts a server at http://127.0.0.1:8080, and opens your browser. Ctrl+C to stop.

The Cowork tree is large (typically 1–3 GB of audit logs and snapshots). First scan walks 500+ JSONLs and may take 60–120 seconds on a heavy user's machine. Subsequent scans are incremental.

CLI

python3 cli.py scan                  # rescan both sources
python3 cli.py scan --code-only      # only ~/.claude/projects/
python3 cli.py scan --cowork-only    # only Cowork
python3 cli.py today                 # today's totals
python3 cli.py stats                 # all-time totals
python3 cli.py tips                  # active suggestions
python3 cli.py dashboard             # scan + serve at :8080

Environment variables

Var	Default	Purpose
CLAUDE_PROJECTS_DIR	~/.claude/projects	Claude Code JSONL root
CLAUDE_COWORK_DIR	~/Library/Application Support/Claude/local-agent-mode-sessions	Cowork session root
TOKEN_DASHBOARD_DB	~/.claude/token-dashboard.db	SQLite cache
PORT	8080	Server port
HOST	127.0.0.1	Bind address. Don't set 0.0.0.0 on networks you don't fully control.

Source toggle

The topbar has a three-way filter (all / code / cowork). Selection persists across reloads via localStorage and is applied to every API call as ?source=.... Cowork projects are prefixed cowork:<workspace-id> so they sort separately in the Projects tab.

Design system

Built using Visionaire Labs' Impeccable skill — a fork of Anthropic's frontend-design skill. Visual identity is documented in PRODUCT.md and DESIGN.md. Edit those before changing the visual layer; both files are loaded by /impeccable commands.

Privacy

Nothing leaves your machine. The dashboard fetches its JSON from 127.0.0.1; CSS and ECharts are vendored. Cmd/Ctrl+B blurs prompt text and other sensitive content for screenshots.

Tech stack

Python 3 (stdlib only) for the CLI, scanners, and HTTP server. SQLite for the local cache. Vanilla JS + ECharts for the UI, no build step.

Data flow: cli.py → scanner.py (Claude Code) and cowork.py (Cowork) → SQLite → server.py exposes /api/* JSON routes and serves web/.

Limitations

• Skill token-per-call is partial when skills live outside ~/.claude/skills/ (inherited from upstream).
• Cowork audit JSONLs include _audit_timestamp instead of timestamp on assistant rows; the scanner backfills, but if Anthropic changes the audit schema this is the first thing to break.
• Pricing math underneath is still pay-per-token rates; the subscription flip just promotes the flat fee to primary.

Contributing

python3 -m unittest discover tests before opening a PR. Stdlib-only.

License

MIT, inherited from upstream. See LICENSE and NOTICE.md for attribution.