Live for Claude Code & Codex

Live token usage tracking for Claude Code and Codex.

A native macOS menu-bar app for live Claude Code and Codex token tracking, AI coding session cost estimates, context usage, and 30-day history. No wrapper, no PATH setup — it hooks straight into Claude Code, so every session shows up whether you launch it from a terminal or your IDE.

macOS 13+ Zero dependencies Apple Silicon & Intel

On Intel? Download for Intel

How it works

Launch Claude. Track everything.

There's no wrapper and no PATH mutation. Tracking runs through Claude Code's own SessionStart / SessionEnd hooks, so every session counts — whether you launch claude from a terminal or from your IDE.

Sessions appear in the menu bar within a second of the first token, and history is finalized even when the app is closed — the hook re-parses the transcript on SessionEnd.

While a session is live, tokenflow tails the usage events written to disk, accumulating input/output tokens, cache reads & writes, per-model cost and context-window usage in real time.

~/projects/tokenflow — zsh
claude
SessionStart hook → tracking session a3f9…
 
Claude Code v1.2 — ready.
> refactor the pricing table
 
↑ 175 ↓ 36.4k cache 8.9k $1.76
sonnet-4.5 · 214 msgs · 1h 37m
 
Everything in the menu bar

Usage you can actually see.

Live stats where you already look, a full per-session breakdown one click away, and 30 days of history grouped by day.

Live menu-bar stats

An always-visible ↑in ↓out readout that counts up with an indigo arrow flash as new usage lands.

Per-session cards

Tool, project, duration, input/output, cost and a per-model breakdown — for every running session at once.

30-day history

A draggable window of finished sessions grouped by day, each expandable to a full cache & cost breakdown.

Cache-aware pricing

Distinguishes cache reads and writes from fresh tokens, with a per-model USD rate table that resolves dated variants.

Terminal & IDE, unified

Hooks into Claude Code itself, so terminal and IDE sessions both show up — no wrapper, no PATH changes. Dragging the app to the Trash cleanly reverses everything.

Context-window meter

Tracks how much of each model's context window a session has consumed, right alongside its cost.

Native macOS 13+ Built in Swift 5.9 Foundation only — zero deps No schema drift, shared Codable
Install

Running in under a minute.

Download the app, or build from source if you'd rather. The menu-bar app handles the rest.

01

Get the app

Download the DMG, open it, and drag TokenFlow.app into Applications.

open TokenFlow-1.1.6-arm64.dmg
02

First launch

Right-click → Open the first time. The app registers a session hook in Claude Code automatically — no shell config, no PATH changes.

hook → ~/.claude/settings.json
03

Just work

Run your tools as usual. tokenflow follows along and the menu bar comes alive.

claude --resume

Stop guessing what a session costs.

Download tokenflow and watch every token — and every dollar — flow by in real time.

macOS 13 Ventura or later · Apple Silicon & Intel · No wrapper, no PATH setup

On Intel-Chip? Download for Intel

Cost calculation

The rates behind every dollar.

tokenflow estimates session cost from this per-model USD rate table. All rates are in USD per 1,000,000 tokens.

Created June 25, 2026 · Source pricing pages cached 2026-06

Anthropic Claude

Cached input 10% of input; cache writes 1.25× (5m) / 2× (1h). anthropic.com/pricing

Model idInputOutput
claude-fable-510.0050.00
claude-mythos-510.0050.00
claude-opus-4-85.0025.00
claude-opus-4-75.0025.00
claude-opus-4-65.0025.00
claude-opus-4-55.0025.00
claude-opus-4-115.0075.00
claude-sonnet-4-63.0015.00
claude-sonnet-4-53.0015.00
claude-haiku-4-51.005.00

OpenAI GPT-5 family (Codex)

Cached input is 10% of input across the family. openai.com/api/pricing

Model idInputOutput
gpt-5.55.0030.00
gpt-5.5-pro30.00180.00
gpt-5.42.5015.00
gpt-5.4-mini0.754.50
gpt-5.4-nano0.201.25
gpt-5.4-pro30.00180.00
gpt-5.21.7514.00
gpt-5.2-pro21.00168.00
gpt-5.11.2510.00
gpt-51.2510.00
gpt-5-mini0.252.00
gpt-5-nano0.050.40
gpt-5-pro15.00120.00
gpt-5-codex1.2510.00

DeepSeek

Real cache-hit input is ~2% of input. api-docs.deepseek.com

Model idInputOutput
deepseek-v4-flash0.140.28
deepseek-v4-pro0.4350.87
deepseek-chat → flash0.140.28
deepseek-reasoner → flash0.140.28

Z.ai GLM — vision models

Ids are lowercase; real cache-hit input ~18% of input. docs.z.ai

Model idInputOutput
glm-5v-turbo1.204.00
glm-4.6v0.300.90
glm-4.6v-flashx0.040.40
glm-4.6v-flashFreeFree
glm-4.5v0.601.80
glm-ocr0.030.03

Z.ai GLM — text models

Free tiers read as $0.00, not n/a. docs.z.ai

Model idInputOutput
glm-5.21.404.40
glm-5.11.404.40
glm-5-turbo1.204.00
glm-51.003.20
glm-4.70.602.20
glm-4.7-flashx0.070.40
glm-4.7-flashFreeFree
glm-4.60.602.20
glm-4.50.602.20
glm-4.5-x2.208.90
glm-4.5-air0.201.10
glm-4.5-airx1.104.50
glm-4.5-flashFreeFree
glm-4-32b-0414-128k0.100.10

Unknown models cost nothing and show n/a rather than $0.00. Date-suffixed ids resolve to their base rate — e.g. claude-sonnet-4-5-20250929claude-sonnet-4-5.