Token Bar

A native macOS menu bar app for monitoring AI coding assistant token usage at a glance.

Features

• Multi-Provider Support — Toggle between Claude Code, OpenCode, or combined view
• Multi-Device — Aggregates Claude Code stats from multiple machines via SSH over Tailscale/LAN
• Local + Remote — Reads OpenCode data from this machine and remote machines via SSH, merged automatically
• Live Token Counting — Scans active session files for real-time today stats, not just cached data
• Activity Heatmap — 16-week GitHub-style calendar with hover tooltips showing per-day metrics
• Cost Estimation — Calculates approximate USD spend per model using API pricing (click for full breakdown)
• Model Breakdown — All-time token usage by model with relative progress bars and cost per model
• Streak Counter — Consecutive days with coding activity
• Peak Hours — 24-hour session distribution chart highlighting your most productive time
• Launch at Login — Toggle to start automatically on boot
• Dark Theme — Designed to look clean alongside your menu bar

Requirements

• macOS 14+
• Swift 5.9+
• Claude Code installed (~/.claude/ directory)
• For OpenCode: Python 3 (local and/or on remote machine)

Quick Start

git clone https://github.com/dougchansan/token-bar.git
cd token-bar/TokenBar
./build.sh
open build/TokenBar.app

The app runs in the menu bar only (no dock icon). Click the ◆ Model | Tokens label to open the panel.

Setup

Claude Code (works out of the box)

If you have Claude Code installed, Token Bar reads from ~/.claude/stats-cache.json and scans session JSONL files under ~/.claude/projects/ automatically. No configuration needed.

• Models: Opus, Sonnet, Haiku (all versions)
• Data: polled every 30 seconds, live session scanning for real-time counts

OpenCode (local)

If you use OpenCode on this Mac, Token Bar reads it automatically too.

1. Install OpenCode (if not already installed)

npm install -g opencode-ai

2. Use it with any provider (Ollama, Moonshot/Kimi, etc.)

# Example: run with a local Ollama model
ollama pull glm4
opencode --model ollama/glm4

# Example: run with a built-in free model
opencode --model opencode/kimi-k2.5-free

3. Token Bar picks it up automatically

The app reads from ~/.local/share/opencode/ — supports both SQLite (v1.2+) and legacy JSON storage (v1.1.x). Click Open in the header to see your OpenCode stats.

OpenCode (remote machine via SSH)

To also pull OpenCode stats from a remote machine (e.g., a Windows PC, Linux server, or GPU rig):

1. Configure the remote host

Edit TokenBar/TokenBar/OpenCodeReader.swift and update the defaults in init():

init(host: String = "YOUR_IP_OR_HOSTNAME",
     user: String = "YOUR_USERNAME",
     scriptPath: String = "opencode-stats.py")

2. Set up SSH key auth (passwordless login required)

# Generate a key if you don't have one
ssh-keygen -t ed25519

# Copy it to the remote machine
ssh-copy-id user@your-host

# Verify it works
ssh user@your-host "echo connected"

3. Deploy the aggregation script

scp scripts/opencode-stats.py user@your-host:opencode-stats.py

4. Test it

ssh user@your-host "python opencode-stats.py"

You should see JSON output with token usage. If you see {"error": "no opencode data"}, OpenCode hasn't been used on that machine yet.

5. Rebuild and launch

cd TokenBar
./build.sh
open build/TokenBar.app

Local and remote data are merged automatically. The status bar shows "Local + 10.0.0.50" when both sources are connected.

Note: The aggregation script (scripts/opencode-stats.py) supports both SQLite (OpenCode v1.2+) and legacy JSON (v1.1.x) storage. It detects the format automatically. The remote machine needs Python 3.

Disabling remote / local

• No remote machine? Leave the default host — if SSH fails, the app gracefully falls back to local-only data.
• No local OpenCode? That's fine too — remote-only works. Install OpenCode only if you want to track local usage.

Provider Toggle

Click Claude / Open / All in the header to switch views:

• Claude — Local Claude Code stats only
• Open — OpenCode stats (local + remote merged)
• All — Everything combined from all sources

Menu Bar Display

Shows your most-used model today and total token count:

◆ Opus 4.6 | 783.9K

When set to OpenCode or All:

◆ Kimi K2 Turbo | 98.5M
◆ All | 99.3M

Cost Calculation

Estimated costs are computed from token counts using published API pricing:

Anthropic (Claude Code)

	Input	Output	Cache Read	Cache Write
Opus	$15/M	$75/M	$1.50/M	$3.75/M
Sonnet	$3/M	$15/M	$0.30/M	$0.75/M
Haiku	$0.80/M	$4/M	$0.08/M	$0.20/M

Moonshot AI (OpenCode)

	Input	Output	Cache Read
Kimi K2 Turbo	$1.15/M	$8.00/M	$0.15/M
Kimi K2.5	$0.60/M	$3.00/M	$0.10/M

Local Ollama models (GLM-4, Qwen, Llama, Mistral, etc.) show $0 — they're free.

Click the cost pill in the app for a per-model line-by-line breakdown.

Architecture

• SwiftUI with MenuBarExtra (.window style)
• No external dependencies — pure Swift + AppKit/SwiftUI
• Polls stats-cache.json every 30 seconds (skips re-parse if file unchanged)
• Scans session JSONL files for live today metrics
• Python script for OpenCode data aggregation (local + SSH)
• Merges local and remote OpenCode data automatically
• Background-only app (LSUIElement = true)

License

MIT