View Source

Hardware-Free LLM Monitoring

A Windows system tray companion that tracks AntiGravity + Cursor token usage from local databases, comparing cooling overheads to human biology.

Dillon Simeone Current Token Usage

This page loads token usage from the tracker cache JSON. Click a period card to view cooling water equivalents.

Loading live data…

Today
Week
Month
Lifetime
1.12 L Water consumed for Today's tokens
44.8% Equivalent of human daily hydration limit (~2.5L)

Model Token Breakdown

Token Usage Trends

Interactive daily breakdown of LLM tokens consumed over the past two weeks.

Delta Scan Execution Protocol

To optimize resource utilization and prevent locking system writes, the background token tracker follows a strict delta-caching execution sequence:

1. Load Cache Log

At startup, the service attempts to read cached entries from token_tracker_cache.json. If found, it populates the local entries list immediately.

2. Compare mtime Handshakes

Before launching a filesystem scan, it queries the SQLite and PB modification times. Files that have not been modified since the last recorded scan are bypassed.

3. Read/Decode Diff Only

Only modified conversation logs are processed. The varint protobuf parser decodes the newly added rows, saving the updated list to the JSON cache log.

Water Footprint Methodology & Citations

Data center cooling footprint estimates are calculated using Google's published environmental data. Google reports that its data centers consume an average of 0.26 mL of water per median Gemini text prompt.

Accounting for average prompt sizes of 500 to 2,000 tokens (including system instructions, contextual window, and outputs), this translates to approximately 0.0003 mL of water per token (or 0.3 mL per 1K tokens). This value accounts for both direct evaporative cooling water on-site and indirect water consumption from power-generation plants supplying the grid.

Python Infrastructure & Integration

1. Pure-Python Protobuf Decoder

The application implements a custom Varint and Protobuf wire-format parser (`proto_decoder.py`) from scratch. It reads length-delimited wire tags (wire type 2), fixed width types (wire types 1 & 5), and 7-bit continuation bit varints (wire type 0). This avoids installing bloated Google Protobuf library wrappers on your system.

2. Delta Caching SQLite Reader

Instead of performing database locks or locking writing handles, `token_parser.py` loads database files in read-only mode using sqlite3 connection URIs (`?mode=ro`). The delta caching system records previously parsed conversation ids and sizes, only scanning databases modified since the last saved scan timestamp.

3. PIL Tray Icon Engine

The system tray process utilizes `pystray._win32` as the backend window wrapper. PIL draws the custom rounded rectangle background, resolves the current token count string (e.g. `1.5M`), maps the color gradient dynamically based on daily target thresholds, and renders the 32x32 pixel tray canvas.

How It Fits Together

01

FileSystem Listener

Wakes up on Startup or once per hour, scanning database directories.

02

Delta Parser

Loads cache log, queries databases for new idx rows, and decodes binary schemas.

03

Dynamic Render

Calculates today's token tally, renders a custom PNG, and notifies the Windows shell tray.