Seeing Your AI Subscriptions: Why I Built a Private Usage Meter

Real Meter interface from the published October 2026 tour; historical snapshot.

The Problem Was a Lid, Not a Model

For a while my most important AI infrastructure was an open laptop. I ran several agents on parallel projects, and I worried that closing the MacBook would interrupt work in flight. So I carried it around open. That is a fragile setup, and I knew it.

The fix was boring hardware. I moved background workloads to an M3 Ultra Mac Studio that I control remotely from the MacBook. The agents keep running; I manage them from wherever I am. That solved where the work runs. It did not solve the question that came next: with roughly $600 a month going to AI subscriptions (my own estimate, not an audited bill), how do I know what each account still has left?

Why Provider Dashboards Do Not Answer It

Each provider shows you something about its own plan. None of them shows you the whole picture. Three different questions get mixed together:

• Account allowance. How much of this subscription's limit remains, and when does it reset? That belongs to the account, whatever device you used.
• Local activity. Which conversations, models and projects did this particular machine run? That belongs to the machine, not the account.
• Intent. Which project should use the next hour of capacity? Neither of the first two tells you.

When a team or a solo builder has several paid accounts, the failure is quiet. One allowance runs out mid-task while another sits unused until it resets. Nothing breaks loudly; you just pay for capacity that expired.

What I Built

Meter is my private dashboard for exactly this. It shows provider-wide remaining limits and reset times per account next to local activity grouped by account, model and project. A "use next" ranking is a planning aid: it looks at remaining allowance, reset time and observed pace, and suggests where work could go.

Some boundaries matter, and I would want a vendor to state them too:

• Meter is a private personal tool, not a product on sale and not a customer deployment.
• Account quotas come from the subscription account. Conversation, token and activity history is scoped to the Mac that Meter is connected to. It does not see other devices or cloud-only chats.
• The ranking does not schedule anything. It does not guarantee that remaining allowance can be used before a reset.
• It does not claim to reduce cost. It makes the spend visible, which is a precondition for reducing it.

The 24-second tour on the original LinkedIn post is a real capture of the production UI, which shows the interface and nothing about provider performance or savings.

A Checklist for Your Own AI Spend

You do not need to build a dashboard. You do need answers to these before AI spend grows:

1. Who owns each account? Shared logins hide who consumes what.
2. What is the unit of limit? Messages, tokens, a rolling window and a weekly cap behave differently.
3. When does each allowance reset? Write the dates down. Expiry is where money is lost.
4. What counts as work? Tag activity by project so a heavy day maps to an outcome, not just a number.
5. Which spend is included and which is metered? Some plans bundle a separate API allowance; our credits explainer covers one such case and its conditions.
6. Where does the work physically run? A machine that sleeps, restarts or runs out of memory is a hidden limit. Mine ran out of application memory even at 96 GB, which I describe in the Mac Studio write-up.

What I Would Not Conclude

I would not conclude that a dashboard saves money. The ledger only tells you where you stand; savings come from decisions you make with it. I would also not treat the $600 figure as typical. It is one builder's spend across several tools, and your numbers will differ.

If you are deciding how to split work between a strong lead model and cheaper helpers, the question of what to measure is the same one: cost and time per accepted result, including retries. The Haiku 5.5 subagent piece applies that to a proposed lead-and-worker test.

Read Next

The original post is on LinkedIn. For Apple's own description of the hardware line, see its Mac Studio announcement; it is product context, not evidence about my workloads. The companion article on Agent, Tracking AI usage across accounts, covers how the same idea works as an operating practice rather than a product story.

If your team is spending across several AI tools and wants someone to set up usage governance and decision rules, that is the kind of work our fractional Head of AI service covers. Otherwise, the checklist above is yours to use. To talk it through, you can get in touch.