Twenty-one questions engineers, leads, finance and admins bring to their AI spend, each with the screen that answers it, who can see it, and which plan holds it.
Cost by provider and model for you, for your team and for the company, month by month and by project, with billed and settled figures side by side.
Each person sees their own spend, a team lead sees the team, and leadership sees teams as totals.
The same vendor now sells you a seat and then bills usage on top: credits, on-demand requests, overage charged after the month closes. A headcount times a seat price leaves out whatever that usage came to.
ai-cur files every charge under how it was bought, as the source recorded it, and keeps seats and usage apart all the way to the total.
Anthropic, Cursor and GitHub each include usage reporting with the plan you already pay for, and it keeps improving. What a vendor's own dashboard is not built to show you is its competitors on the same page, in one currency, against the same people and the same merged work. That page is the product.
| Vendor | Cost per person | What the cost figure is |
|---|---|---|
| Anthropic, Claude Code | Yes | Estimates, labelled as such; the billing page is the authority1 |
| OpenAI | Not by person | Billed amounts, grouped by project, line item or API key2 |
| Cursor | Yes | Spend per user, through the Admin API3 |
| GitHub Copilot | No cost field | Usage metrics that carry no cost, in the API schema we read4 |
Four vendors, four different cost models: two of the four report cost per person, and no two report it the same way. From each vendor's own documentation, read on 8 August 2026.
Each paid seat is classed as active, idle, not attributable or not measurable for the month, and the classes add up to the whole seat bill. Idle means the vendor reported usage that month and none of it was the seat holder's.
In the demo, August's idle seats cost 280.00 USD.
Copilot, Cursor or Claude used by people with measured activity, on a vendor you do buy seats for, where the billing ai-cur reads holds no seat for them.
That seat may sit on a reseller's invoice or a parent company's agreement, so the page says no seat is recorded here, and never that one is missing.
Plan fit sets each seat's price against what the same measured usage would cost at list prices. The list figure is an estimate that sits beside the seat cost and is never added to it.
No usage record carries a list price by itself, so until your admin declares a rate card the answer reads Cannot say.
Cost per merged pull request, per resolved issue and per epic, each printed with the cost and the count it was divided from and the share of people it covers. Seat purchases stay out of the cost, because a seat is capacity, not work.
Team figures wait for a minimum cohort, so a very small team never has one person's month published as its own.
A lead sees their own team in full and peer teams only as totals, with nobody outside their team named.
A comparison appears only when a team's roster meets the minimum your admin declares. Below it the row reads Unavailable, and its cost still counts in the company total.
Settled cost is split across the projects each person resolved Jira issues in that month, and the column adds up to the company total.
No usage record names a project, so the split follows work references and is stated as a share, never as what a project consumed. Cost with no project gets its own row, at full size.
Declare a monthly budget for a team, a project or a person, with alert thresholds and soft and hard caps. ai-cur works out the cap state and serves it as a decision your own gateway or key issuer can act on.
It ships the decision, not a gateway, so nothing here blocks a request by itself.
The finance view closes a month by cost centre, provider and commercial model, and holds open charges back until the vendor settles them. No individual is named anywhere in it.
The chargeback export is one CSV per month and currency, built from the same settled rows, so the file's total is the page's total.
Declare the commitments in the contracts you hold, and ai-cur tracks each drawdown from settled charges only, in the currency it was declared in.
Each charge is marked by its rate basis: a contracted price recorded, a list price, or neither, which stays unknown instead of being guessed.
Every amount carries its currency, and no total anywhere adds two together. Where a group holds more than one currency, its total is withheld and the page names the currencies it found.
A metric takes the currency of its cost. Nothing is converted, defaulted or assumed.
Each source carries an access state: never attempted, disabled, denied, failed, stale, live or unscheduled. A page names its incomplete sources right above the figures they touch, so a missing feed never passes as a silent zero.
A correction is a new row that names the charge it supersedes. The original is never edited or deleted and stays visible, and a restated figure carries a restatement mark on the page.
Every served figure keeps its lineage back to the raw charge it came from.
Quota and credits are the free Tracker's job, on your own Mac: it reads your Claude, Codex and Cursor subscription gauges, shows paid credit state, and warns before a limit runs out.
The hosted product does not track provider credit balances.
As teammates install. Every install gets a token of its own, tied to the person who runs it, and once people are mapped to a team the lead's view fills in, member by member.
People without a team and installs without a person are listed too, so nobody's cost drops out of the totals.
Invite teammates by email and a personal account turns into a team; becoming a team keeps everything the person already had. Members hold member, admin or owner roles, with a single owner.
Inside the report, admins declare management, admin and finance roles, and the database decides which rows each role can read.
Teams sign in with a magic link by email, or with Google or Microsoft. Business customers sign in through their own identity provider, such as Okta, and team membership comes from the Okta groups an admin maps, never from who happened to sign in.
Sign up and Sign in are the two doors in the header. The same page sends a link by email or continues with Google or Microsoft; there is no password to keep.
Yes. A team runs a Team trial with no card, and when the trial ends without one it simply ends; nothing is charged. There is one trial per team.
The trial starts from your team's billing page after sign-up, through Stripe checkout, and the same portal is where a card is added or the plan is stopped. The steps are on the trial page.
Download an export of what is held about you whenever you like. Deleting your account starts a governed deletion: a link by email lets you cancel it for a short while, and the purge itself waits out a grace period before anything is erased.
The hosted service signs an entitlements document for your account, and the Tracker checks that signature against a public key set anyone can fetch. A feature your plan does not include stays locked, and an unknown feature unlocks nothing.
The hosted service has an intake for the Tracker's work report, the per-project and per-task summary it builds on your own machine. The intake stays off until an operator turns it on.
The AI Usage Tracker is a free, open-source menu bar app for macOS. It needs no account, and it is where anyone can start before a team exists.
Claude Code and Codex sessions in the files they already write on your Mac, every minute, and, if you choose, your Claude, Codex and Cursor subscription gauges.
Usage trends, paid credit state, and alerts on quota thresholds, projected exhaustion and stale credentials. An API-only mode skips the subscription gauges.
A local work ledger assigns sessions to projects and tasks, with a review window to correct them and daily or weekly reports. A ratio it cannot divide honestly is left blank, not zero.
The activity timeline is a contentless index: session text stays in the files it came from.
Every cost record ai-cur keeps is shaped after FOCUS, the open billing data specification, and the finance export names its columns in FOCUS terms6. That makes it FOCUS-aligned. No validator has tested it, so it is not a certification.
The record already keeps the model and the input and output tokens separately wherever the vendor reports them, which is where the specification's next version is scoped to go.
Gartner's 2026 evaluation of AI coding agent vendors counts cost controls and usage analytics among the capabilities it expects5. Expect every vendor's own reporting to keep getting better.
That helps you, and it leaves the shape of the problem where it was: each vendor improves the view of its own tools, while your bill, your people and your merged work span all of them.
Sign up, install the Tracker, invite the team. Team runs for sixty days with no card, and ends on its own if you never add one.
