To track Claude Code usage per developer, route Claude Code through a gateway that tags every request to a user and a project. Then the tool records who made each request, which model it used, and what it cost. You read the result as live and historical spend, broken down by developer and by project. That is the whole method, and it takes minutes to set up.
Anthropic's own analytics dashboard shows per-user spend for the month on API billing, which is a good start if you use nothing but Claude Code. But the moment your team also runs Codex, or you want cost grouped by client project instead of just by person, you need one control point that sees everything. A gateway is that point.
The blind spot without per-user, per-project tracking
Most teams learn where their AI money goes only when the invoice arrives, and by then it's one big number you can't break apart. You know the total. You don't know who spent it, on which project, or whether an expensive model did work a cheaper one could have done. The spend is real: Anthropic's docs put Claude Code at roughly $13 per developer per active day, about $150 to $250 per developer per month, with heavy users well above that. Multiply that across a team and a shared key, and attribution becomes guesswork.
This is not a hypothetical problem. When Uber gave Claude Code to about 5,000 engineers, it burned through its entire 2026 AI budget in roughly four months, with cost per engineer reaching $500 to $2,000 a month before it reached for blunt caps. You can't manage what you can't see, and a single line item on a provider bill shows you almost nothing.
What you can actually see
Once every request carries a user and a project tag, the gateway turns the bill into a report you can read. With Clawgate you see:
- Spend per developer:who is running Claude Code, how much they've spent this day, week, and month, live as sessions run.
- Spend per project:cost grouped by project, so you can tell at a glance which project is spending most on AI.
- Which model was used:a per-model breakdown, so you can see when a premium model is doing routine work.
- Live and historical:current-session numbers for right now, plus durable monthly snapshots for trend and reconciliation.
Attribution is automatic because each key is tied to one person and one project. There's no metadata discipline to enforce and no tagging for developers to remember. If a request went through the gateway, it landed in the right bucket.
Why per-project tracking matters for client work and chargeback
If you bill clients, per-project cost isn't a nice-to-have, it's the number you invoice. Set a project per client, run that client's work under it, and the project total is exactly what you charge back. The same logic applies internally: allocate AI costs by team or project so each cost center owns its own usage instead of hiding inside a shared line. When finance asks why the AI bill moved, you can point at the project that moved it, not shrug at a total.
This is also how you keep a productive team productive. Instead of rationing AI across the board, you find the one project or the one workflow that's expensive and fix that. For more on turning this visibility into hard limits, see how to control Claude Code costs and how to set a Claude Code budget that actually stops spend.
What the per-project view tends to reveal
The first thing per-model visibility usually surfaces is model choice. In a small experiment we ran, the same coding task cost about 12 times more on the priciest model than the cheapest, same prompt, six models. Without a per-project, per-model view, a team can be paying that premium on routine work and never know. With it, the expensive project stops being a mystery: you can see the tokens, see the model, and decide whether that project should route to a cheaper capable model or compress its context.
We're keeping this honest, so no invented dashboard numbers here. The point is what the view lets you do: turn a flat bill into a ranked list of where tokens actually went, then act on the top of that list. That is the difference between real AI spend control and a monthly surprise.
Tracking is where Clawgate started
Clawgate began as an internal tool at Virstack. We built the per-user, per-project view first because we needed to know where our own AI bill was going, and that same visibility helped us cut our bill by 70% before Clawgate was ever a product. What we learned building it is that no single existing tool did all three jobs a team actually needs, which is why we built one that does:
- Monitor and control:the per-user and per-project tracking in this guide, plus hard-stop budgets and model control.
- Cut token usage:opt-in prompt compression that cuts up to 92% of the tokens agents resend each turn on heavy workloads, plus routing work to cheaper capable models. The providers won't build this, because tokens are their revenue.
- Catch abuse and impersonation:content-fingerprint detection that flags leaked or shared keys and requests whose content doesn't match the project they claim.
That last point rides on the same tracking. Because every request is attributed, a key that suddenly shows usage from a different project, or spikes in a way its owner wouldn't, is visible instead of buried. So the tracking view isn't only a bill you can read. It's also how you spot a shared or leaked key before it costs you.
How to set it up
The workflow is short. Create a project (one per client or team), issue each developer a vsk_… key bound to their user and that project, and point Claude Code at the gateway with that key. From then on every request is authenticated, attributed, priced, and forwarded to the model, and the dashboard fills in per-user and per-project spend on its own. No local config juggling, no per-developer setup beyond the one key. If you also run Codex, opencode, or Cursor, they point at the same gateway and land in the same reports.
Frequently asked questions
Can I see Claude Code cost per project?
Yes. When Claude Code runs through a gateway that tags each request with a project, cost is grouped by project automatically. In Clawgate, every key is tied to a project, so you get a per-project spend view with no manual tagging. That is how you see which project is spending most on AI.
Does it track usage in real time?
Yes. Because every request passes through the gateway, usage is recorded as it happens. Clawgate shows live spend per user and per project as sessions run, and keeps the historical record so you can compare weeks and months later.
Can I export the usage data?
Yes. Clawgate keeps per-user and per-project usage and cost in the dashboard, and produces monthly invoice snapshots with a per-model and per-user breakdown, so you have a durable record to reconcile against your own accounting or bill back to a client.
Can I charge AI costs back to a client or team?
Yes. Because usage is attributed to a person and a project, you can allocate AI costs by team or project and bill client work back to the right client. Set a project per client, run that client's work under it, and the per-project total is your chargeback number.
Written by Chathuranga K, an engineer at Virstack.
See exactly where your Claude Code tokens go
Connect Claude Code and Codex once, tag every request to a developer and a project, and read live and historical spend broken down by both. Attribute AI cost to the right team, then set a hard cap where you need one.