Standard seat or API budget?
Compare each person's monthly cost at API prices with the price of their seat. When it's consistently lower, an API token budget is the cheaper way to give them Claude Code.
Evidence: estimated cost per person, by month
Actionable intelligence for Claude Code
Pay for the Claude seats your people actually use.
Recon measures how much Claude Code each person really uses: what it would cost at API prices, and how much of their subscription quota they get through. You see who needs a premium seat, who'd be fine on standard, and who'd cost less on an API budget. The data never leaves your infrastructure.
- Per-person, per-model figures from real use, not surveys
- No prompts, responses or file paths leave a laptop
- Runs on your own VM or Kubernetes, for macOS and Windows users
| Est. cost | 5-hour quota | ||
|---|---|---|---|
| Person | at API prices | average | highest |
| priya | $412.60 | 71% | 100% |
| sam | $188.05 | 38% | 64% |
| lee | $46.10 | 9% | 17% |
| kim | $7.85 | 2% | 6% |
- priya hits her limit: the case for a premium seat.
- lee on a premium seat rarely passes 17%: a standard seat would do.
- kim uses $8 a month at API prices: an API budget costs less than a seat.
Seats bought on guesswork
Premium seats cost several times a standard one. Without usage figures you can't tell who's getting the value.
Light users on full seats
Someone who opens Claude Code twice a week may cost far less on pay-as-you-go API usage than on a subscription.
Heavy users held back
People who hit their 5-hour or weekly limit mid-task lose time. You only hear about it when they ask for more.
Decisions it answers
Turn Claude Code usage into seat and budget decisions
Is the premium seat paying for itself?
See how much of their quota premium users actually use, window after window. If the average and the highest stay well within a standard seat's limits, move them down.
Evidence: average and highest quota per completed window
Who needs more?
People whose 5-hour and weekly windows regularly peak at their limit are being slowed down. They're your case for an upgrade, with the figures to back it.
Evidence: quota peaks, live and over time
Which models earn their cost?
Split usage by model to see where the most capable model is used, and where a smaller, cheaper one would do the same job.
Evidence: cost and tokens by person and model
Recon gives you the figures; the call is yours. It doesn't need to know your seat prices, and it never sees what people use Claude for.
What you see
The reports behind every decision
Usage by person
Estimated cost, input, output and cache tokens, and active time per person, across all their devices.
Usage by model
The same figures split by model, so you can see where Opus is worth it and where a smaller model would do.
Subscription quota by window
Each person's live 5-hour and 7-day quota, plus the average and highest of their completed windows: how much of a seat they really use.
Devices and health
Which laptops are sending, and whether your stack is healthy. Revoke a device and it stops at once.
The right view for each role
Admins and viewers see everyone. Members see only their own figures.
An audit trail
Every report request, setting change and sign-in is recorded in a tamper-evident log you can read.
How it works
Collected on your side. Read live. Never stored by Recon.
-
1
The Recon agent
A small agent on each laptop reads Claude Code's own usage metrics and status line, and sends 5-minute totals per model. Nothing per turn.
-
2
Your Recon stack
Runs on your infrastructure, on Docker or Kubernetes. Devices send to it over your network, and it keeps the figures in its own database.
-
3
The Recon portal
Your stack connects out to Recon over an encrypted tunnel. Reports stream through the portal to your browser and aren't kept anywhere along the way.
Setup: a stack per organisation, one enrolment key per laptop, then one command connects Claude Code.
Security and privacy
Built for teams that can't send usage data to someone else's cloud
- Your data stays with you. Usage figures live in a database on your infrastructure. Recon's own systems have nowhere to store them.
- Aggregates only. No prompts, responses, file paths, session IDs or per-turn records are collected. Usage from personal Claude accounts is left out.
- No inbound ports. Your stack connects out to Recon over an encrypted tunnel. Nothing on the internet can reach it.
- You decide on operator access. Recon's operators see your stack's health, not your figures, unless you turn access on. Your stack enforces that itself.
- Signed instructions only. Settings, report requests and updates are signed, and your stack rejects anything else.
- Signed updates with rollback. Your stack checks every image's signature and rolls back an update that isn't healthy.
- Your sign-in. Users sign in through your identity provider (Entra ID, Okta, Google Workspace) or by email link, limited to your verified domains.
- Regional by design. Your organisation's accounts and settings sit in one region, starting with the EU. There's no global directory of customers.
Join the pilot
We're onboarding a small number of organisations. You get the full product, direct access to the team, and a say in what comes next.
FAQ
Questions people ask first
Does Recon tell us which seat each person should have?
It gives you the evidence: each person's cost at API prices, and how much of their quota they use window after window. Recon doesn't know your seat prices or who holds which seat, so the decision stays with you, and the reports make it a quick one.
Can Recon see what our developers type?
No. The agent keeps only totals: estimated cost, token counts, active time and quota, per model, per 5 minutes. Prompts, responses, file paths and session IDs never leave the laptop.
Is the cost figure what we actually pay?
It's Claude Code's own estimate of what the usage would cost at API prices. On a subscription that tells you the value people get and which models drive it. Recon labels it as an estimate everywhere.
Which Claude accounts does it work with?
Usage figures come from any Claude Code account. Subscription quota needs a subscription account: API-key accounts don't report quota, and Recon shows it as unknown rather than zero.
What do we need to run it?
A small Linux VM (2 vCPU, 4 GiB, around 1,000 devices) with Docker, or a Kubernetes namespace. The stack needs outbound access to Recon only, and devices need to reach it on your network or VPN.
Can Recon's staff see our figures?
Only if you turn operator access on, and you can turn it off at any time. Your stack enforces the setting, and every access is audited.
What if someone leaves?
Revoke their devices from the portal and their stack refuses them at once. Their sign-in can be deactivated the same way.