cursortrack

Track Cursor AI usage and estimate cost per git repository — built for freelancers and agencies who need to invoice clients for AI spend.

Cursor doesn't expose per-project costs. cursortrack reads Cursor's local SQLite databases, attributes every conversation to the git repo it belongs to, applies current model pricing, and presents everything in a live browser dashboard with PDF export.

---

Features

• Per-repo cost tracking — every AI conversation attributed to its git repository via commit history, workspace databases, and chat context
• Accurate token data — input tokens read directly from Cursor's local SQLite; output tokens estimated via context-window delta method
• Live dashboard — dark-mode web UI with period filters, project visibility controls, cost breakdown chart, and daily usage chart
• Auto-sync — data re-imported every 15 minutes; model pricing fetched from cursor.com every 24 hours
• PDF export — print-ready cost report with KPIs, charts, and per-project detail cards; respects active filters
• Persistent history — conversations saved to a local SQLite database; data survives chat deletion in Cursor
• Automatic pricing — headless Chromium scrapes the Cursor pricing page and keeps prices.json up to date; historical costs are never retroactively changed

---

Requirements

• Python 3.10+
• Windows (reads %APPDATA%\Cursor — macOS/Linux paths differ but the code can be adapted)
• Cursor installed and used at least once

---

Installation

git clone https://github.com/jeremypot/cursortrack.git
cd cursortrack
pip install -r requirements.txt
python -m playwright install chromium

pip install -r requirements.txt installs playwright. The second command downloads the Chromium browser binary (one-time, ~170 MB) — this step can't be automated via pip. Playwright is only used to scrape model pricing from cursor.com; everything else uses the Python standard library.

---

Quick start

python dashboard.py

Opens the dashboard at http://localhost:8765, runs an initial data sync, and starts auto-syncing in the background.

python dashboard.py --no-refresh   # skip initial sync, show last cached data
python dashboard.py --port 9000    # use a different port

Command-line only

# Sync all data and write cursor-usage.json
python tracker.py

# Filter by date range
python tracker.py --last 30
python tracker.py --since 2026-01
python tracker.py --from 2026-01-01 --until 2026-03-31

# Update prices only (no data sync)
python tracker.py --prices-only

# Import actual billed costs from a Cursor usage CSV export
python tracker.py --billing-csv "C:\path\to\usage-events-2026-05-26.csv"
python tracker.py --billing-csv usage-events.csv --last 30

Actual billed costs (CSV import)

Import is required for accurate costs. Local estimates miss cache reads and often mis-price models (e.g. Opus); totals can be 10× too low.

1. Export usage-events-*.csv from cursor.com/settings (Usage → Export).
2. In the dashboard, click Import CSV (header or yellow banner), or run:

   python tracker.py --billing-csv "path\to\usage-events-2026-05-26.csv" --last 7

Cursor's CSV has no project column. cursortrack attributes each billed row to a repo via timestamp match (±30s) to local conversations, then proportional token weighting for unmatched rows. The report adds actual_cost_usd alongside estimates; totals match the CSV.

---

Dashboard

Feature	Description
Period filter	All time · Last 7/30/90 days · This month · Custom date range
Project filter	Hide/show individual projects — persisted across sessions, excluded from all totals and charts
Summary KPIs	Total cost · Conversations · Requests · Input tokens · Output tokens · Files edited
Cost breakdown	Horizontal bar chart per repo
Daily usage chart	Stacked bar chart — cost per day per project
Project cards	Per-repo detail: cost, requests, models used, attribution method, date range
Rates dropdown	All current model rates, sync schedule, and Update Prices button
Auto sync	Data re-imported every 15 min · Prices re-fetched every 24 h
Export PDF	Print-ready report — summary, charts, and project cards; projects section starts on a new page

---

How it works

Attribution

Every conversation is attributed to a git repository using three layers, checked in priority order:

Priority	Source	What it covers
1 — commit-linked	~/.cursor/ai-tracking/ai-code-tracking.db	Every file the agent touched → git root. Most accurate.
2 — workspace-sqlite	%APPDATA%\Cursor\User\workspaceStorage\*/state.vscdb	Conversation ID → workspace folder → git root. Covers all single-folder workspaces.
3 — bubble context	%APPDATA%\Cursor\User\globalStorage\state.vscdb	Files referenced in chat bubbles. Good fallback for pure-chat sessions.

Cost calculation

estimated_cost = (input_tokens  / 1,000,000 × input_rate)
               + (output_tokens / 1,000,000 × output_rate)

Input tokens are read directly from contextWindowStatusAtCreation.tokensUsed in Cursor's local SQLite — this is the exact context window size sent to the model.

Output tokens are estimated using the delta method: the growth in tokensUsed between consecutive requests within a conversation approximates what the model generated. Negative deltas (context-window compression) are discarded. Single-request sessions fall back to the flat _avg_output_tokens value in prices.json (default: 2000).

The delta method includes thinking tokens for models like claude-*-thinking, which can inflate the output estimate. It is a conservative upper bound on true output cost.

Pricing

Model rates are stored in prices.json and fetched automatically from cursor.com/docs/models-and-pricing every 24 hours using a headless Chromium browser (Playwright). The "Show more models" button is clicked automatically to capture all models.

• Price changes apply only to conversations synced after the update — stored costs in history.db are never retroactively recalculated.
• prices.json can be edited manually to override any rate.
• The _avg_output_tokens key controls the flat output fallback for single-request sessions.

Data persistence

All conversation data is written to history.db (local SQLite, gitignored). Syncs are incremental — a conversation is only re-written when its token counts have increased (i.e. new messages were added). Deleting a chat in Cursor does not affect history.db.

---

Files

File	Purpose
dashboard.py	Local HTTP server + browser dashboard
tracker.py	Data sync engine and CLI report generator
prices.json	Model rates — auto-fetched; edit to override
requirements.txt	Python dependencies (playwright)
history.db	Conversation history (auto-generated, gitignored)
cursor-usage.json	Latest report snapshot (auto-generated, gitignored)

---

Why not use the Cursor API?

cursor.com/api/usage only tracks the legacy per-token gpt-4 counter. For Ultra and other flat-rate plans it returns zeros — Cursor does not expose per-request token detail through any public API. cursortrack reads exclusively from local SQLite databases.

---

Contributing

Pull requests welcome. A few notes:

• The project has no runtime dependencies beyond playwright for pricing scrapes — keep it that way.
• tracker.py is pure data; dashboard.py is pure presentation. Keep them separate.
• Pricing changes must never retroactively update stored costs (see sync_to_history() in tracker.py).

---

License

MIT