A meter for agent spend · Mac · iPhone · Watch

Know whether you can finish, before the agent stops.

MeterTab reads the logs your AI tools already write and turns them into one reading: how much of your window is gone, how fast it is going, and when it comes back. It is the only usage meter your agents can query mid-run - and the only one that will refuse a tool call on your behalf.

Download on the App Store How agents use it
$19.99once, for all three
platforms
9providers read from
local logs
0servers, accounts
or analytics
Reads
Claude Code Anthropic API OpenAI OpenRouter Codex Cursor Gemini Groq Together AI
The design rule

A coloured pixel means trouble. Nothing else.

Most meters are green when you are fine. That trains you to ignore them, and it wastes the one signal you actually need. MeterTab is grey when you are fine - so the moment anything is coloured, it is worth looking at. Every state is also a shape and a word, so it survives greyscale, Increase Contrast, and colour blindness.

Nominal
square
63%used

The window resets before this pace could exhaust it. No action.

Tight
triangle
68%≈96% at reset

The hatched span is the projection at this pace. The tick is where the clock saves you - just.

Critical
circle
95%8 min to reset

The panel re-ranks itself: this reading takes the full width and the rest compress away.

Amber and red are the only two hues in the product. Charts get a whisper of colour for identity; nothing else is tinted at all.

Why it is different

Your agent can ask how much is left - and be told no.

MeterTab ships an MCP server inside the app, over stdio: no port, no socket, no key to leak. Point a coding agent at it and it can read your live budget mid-run, then downshift the model or pause before it walks into the wall.

get_budget

How much of the window is left, right now.

should_continue

A yes or no the agent can branch on without parsing anything.

get_service_status

Is it your budget, or is Claude itself degraded?

PreToolUse hook

The only part that enforces: it refuses the held tool call when a ceiling is reached, and keeps refusing while it stands.

Not a mockup

This is the app, running on a Mac.

One dropdown from the menu bar: every provider, both windows, burn rate, and where each number came from. No title bar, no sidebar, no scrolling - 500 pt wide and as tall as it needs to be. The figures are MeterTab’s own sample data, labelled as such in the app itself, because showing you a stranger’s real usage would be its own small dishonesty.

Session and week as co-equal constraints, each with its own anatomy - you can tell them apart with the labels covered
Burn rate that projects forward to the moment the limit lands, if the current pace holds
Connect your Claude account and the figures are your account’s own, not an estimate
Ten languages, light and dark, Increase Contrast and full VoiceOver on the gauge
The MeterTab menu-bar panel: a Nominal status strip, then session, week, burn rate and plan value as four equal readings, each with where its number came from, and Claude, Codex and Cursor listed underneath with token counts.

Unretouched · v1.4 · sample data, not real usage

What you get

Specification

Every row is something the app does today. Nothing here is a roadmap item.

01

Answer your agent from anywhere

When Claude Code asks whether a tool may run, MeterTab holds the call and asks you - on the Mac, iPhone or Watch, behind Face ID if you want it. Multiple-choice questions get answered from your phone too. Ignore it and the prompt appears in your terminal exactly as it always did, so it can only ever help.

02

Budget ceilings that actually stop it

Set ceilings per session, day, week or single request - in tokens, with the API-equivalent shown beside them. When one is reached the hook refuses tool calls and keeps refusing, which is what makes a runaway agent give up. Approving instead grants runway for a stretch of time, not one call.

03

See what every session is doing

Which worktrees are busy, the command in flight, how full each context is, and the agent’s own task list. Open a row for the last ten tool calls in full - the thing a truncated log never shows you.

04

Ninety days of history

A heatmap shaded by percentile rather than by your biggest day, so one 400M-token Tuesday doesn’t flatten the other eighty-nine into nothing. A day you didn’t work and a day it couldn’t see are drawn differently, and both are named in the legend.

05

Cost per model

Where the tokens go, by model and provider. On a subscription the dollars are API-equivalent - what the same usage would have cost without your plan - and they are labelled that way everywhere they appear, never as a bill.

06

Knows when Claude is down

MeterTab watches Anthropic’s status page, so “why is this failing?” gets an answer that isn’t always “you’re out of tokens.” Show the indicator during outages only, always, or never. Your agents can ask as well.

07

Exact billing, optional

Add your own provider key and MeterTab pulls real dollar spend from Anthropic, OpenAI and OpenRouter. That is the one place in the app where a bare dollar figure appears, because it is the one place the money is real.

08

Local-first, no account

No backend, no sign-up, no analytics. Keys stay in the Keychain, this-device-only. Your prompts, your code and your file contents never leave the Mac. What your iPhone sees is opt-in, goes to your own iCloud, and is scrubbed of anything shaped like a credential. Data Not Collected.

The part nobody else does

Every number tells you how much to trust it.

Anthropic does not publish a token limit, so any tool claiming an exact percentage is either reading your account or guessing. MeterTab does both - and says which, in the same glance as the number. An estimate never wears the confidence of a measurement.

Measured

Read directly from your account, seconds ago.

Official

A published figure, such as the weekly reset.

Estimated

Derived, or learned from your own history.

Interpolated

The span between two exact readings, marked on the bar itself.

On a subscription
API-equiv≈$7,881.46/mo

Notional value - your tokens priced at pay-as-you-go rates. It is what the plan saved you, not a bill. The qualifier is part of the component, so the figure cannot be rendered, copied or truncated without it.

With your own API key
$412.68
official · billing API

Money actually billed, pulled from the provider’s own billing API. Bare type, no pill, no “≈” - a different object on the page, so you never have to read a caption to know which of the two you are looking at.

MeterTab on iPhone: Claude service status, then the Nominal cluster with the 5-hour session at 62% used, its reset countdown, the weekly bar and burn rate.
MeterTab on Apple Watch: the 5-hour ring at 62% used, time to reset, and the weekly bar underneath.
One purchase, three platforms

On your wrist before you open the laptop.

Pro unlocks native iPhone and Watch apps, home-screen and Lock Screen widgets, and a complication with a live reset countdown - all syncing your Mac’s numbers over your own iCloud. The language degrades by dropping layers, never by restyling: the phone keeps the whole gauge, the complication keeps one number and a state.

Approve or deny a waiting call from the notification
Optional Face ID before you allow
Pick an answer to a multiple-choice question
The live session list in your pocket
Widgets configurable by metric
One small summary syncs, not your history
Pricing

Pay once. Never a subscription.

A tool that tells you what you are spending should not quietly be part of it. The gauge is free forever; everything else is one Universal Purchase.

Free

$0forever
Live usage gauge
Up to 4 local providers
Threshold notifications
Reset countdown
Menu-bar glance
Start free

Pro

$19.99once
Mac + iPhone + Watch, same Apple ID
Universal purchase
Everything in Free
Unlimited providers
The MCP server for agents
Guardrails that refuse calls
Exact billing via your keys
iPhone & Watch apps
Answer prompts from any device
90-day heatmap & burn velocity
Get Pro 14-day free trial
no card
Before you decide

Honest comparisons, including the rows we lose.

This category has good free tools in it. Pretending otherwise would be the same dishonesty the app refuses.

vs ccusage

The reference CLI - free, MIT, and covering more agent CLIs than we do.

vs Usage for Claude

The closest App Store rival. One purchase versus tiers; Claude-only versus nine providers.

Build the hook yourself

Fifteen lines that refuse a tool call when your budget is gone. Free, and tested before publishing.

The weekly limit, explained

What Anthropic documents, what it doesn’t, and where your real reset time lives.
Questions

Answered plainly.

Something missing? [email protected]

How does it know my usage without an account?
+

It reads the log files your AI CLIs already write on your Mac - Claude Code, Codex, Cursor. You grant a folder once; access is read-only and nothing leaves the device. For exact dollar figures you can optionally add a provider API key, which is stored in the Keychain.

Anthropic doesn’t publish a token limit. How can it know how much is left?
+

Three ways, best first. Connect your Claude account and the 5-hour and weekly figures are your account’s own numbers. While connected, MeterTab also learns your personal limit by watching how far your usage moves that official percentage - so the gauge stays accurate after you sign out, and on your iPhone and Watch. Prefer not to connect? Type the percentage from Claude’s usage screen in once, or let it learn the first time you hit a limit. If some of your usage happens on claude.ai or another machine, MeterTab says so rather than pretending the number is exact.

Are the dollar figures money I actually spent?
+

Only if you have connected a provider API key - that is metered spend, and it is real. On a Claude subscription the dollars are API-equivalent: your tokens priced at pay-as-you-go rates, i.e. what the same usage would have cost without a plan. They are labelled that way everywhere they appear, and shown against what you actually pay so you can judge whether the plan is worth it.

What exactly can it stop?
+

MeterTab holds no process and cannot kill your agent. What it has is a hook inside the tool-call loop, so it can refuse the call in front of it - and keep refusing every subsequent one while a ceiling stands, which is what makes an agent give up in practice. Approving works the other way: it stands aside for a stretch of time rather than for one call, because agents fire calls seconds apart.

Does it tell me when Claude itself is broken?
+

Yes. It checks Anthropic’s public status page every few minutes and can show it in the menu bar during outages only, always, or never. Open incidents and scheduled maintenance appear in the app in Anthropic’s own wording, with a notification when one starts, updates or clears. Your agents can ask too, so they stop retrying a service that is down.

Do my keys or usage ever leave my Mac?
+

No. There is no MeterTab backend. Keys live in the Keychain, this-device-only. The only thing that syncs to your iPhone and Watch is token counts and percentages, over your private iCloud - never keys, never message content, and scrubbed of anything shaped like a credential.

Is it really a one-time purchase?
+

Yes. One non-consumable in-app purchase, no subscription, and it is a Universal Purchase - buy once and it unlocks on Mac, iPhone and Watch under the same Apple ID. There is a 14-day free trial and no card required to start.

What are the requirements?
+

macOS 14 Sonoma or later for the Mac app; iOS 17 and watchOS 10 for the companions. Universal binary, Apple Silicon and Intel. Localised into ten languages.

Stop finding out when the agent stops.

Free forever for the gauge. $19.99 once for the rest, across all three platforms.

Download on the App Store Support