Cardinal by RAMCOREOpen Cardinal

Your agents

AI agents and the runner. Eight agents watch your numbers and suggest what to do: seats, coaching, scope, follow-ups, cash, client updates, lost deals and similar past work. Cardinal AI answers questions. Nothing is sent without your click.

Each agent works the same way. Code looks at your numbers at most once an hour and decides whether anything needs you. If it does, you get an item with the facts and, where it helps, a draft. You decide what happens next. Claude only adds words, and only when you click "Word it with Claude": the job runs on your own computer, with your own Claude plan, through the Cardinal runner.

The rules every agent follows

  • Every number is computed by Cardinal. Claude may only point at a number by its key; Cardinal fills it in, and rejects any wording with a typed amount, percentage or hour count.
  • Nothing is sent without your click. A draft has Copy and "Mark as sent": you can send it yourself from your own email or WhatsApp. Execs can also use "Send via comms" on a follow-up or status update: it records exactly what would go, and sends it by email once the workspace sends for real (see "Comms" and "Email sending"). Until then it is a dry run.
  • You see only your own items. Coaching tips are private to you; your manager sees your numbers, never your tips.
  • Seat, cash and deal items are for execs and the workspace owner only, the people who already see those numbers.
  • Cardinal AI can accept, snooze, dismiss or mark done items about an invoice, project or deal for you, as their card offers, with Undo; others you decide here.

The agents

  • Seat optimiser (Team page): seats unused for 30 days, seats used under 40% of the next plan down for two months with no rate limits, and seats rate limited on 4 or more days a month. It shows the monthly saving in rands and drafts a note to the person. An upgrade also says how long the person was blocked and how many blocked hours a month pay for it; a downgrade says how the busiest day compares with the smaller plan. You change the plan at Anthropic.
  • AI coach (Me page): up to three tips from your own last 7 days of Claude Code use, from the waste patterns and your way of working.
  • Scope guard (Deals, and a project's page): a delivering deal whose cost passes 60%, 85% or 100% of its value, and a project budget past 75%, 90% or 100%. It drafts a stop-and-send message to the client that names no internal figure.
  • Follow-up nudger (Deals and Invoices): a deal before delivery with no activity for 7 days, and an invoice past its due date. It drafts the follow-up or the reminder.
  • Cash-flow forecaster (Invoices, workspace owner): your bank balance for the next 13 weeks from your ledger (or the balance you type), invoices still owing on their due date plus how late each client usually pays, and your monthly costs. It warns when the balance drops under your minimum.
  • Client status update (Deals): from Friday 14:00, a weekly email per delivering deal from the hours tracked on it.
  • Loss post-mortem (Deals): asks why each lost deal was lost. Every fifth reason, it counts the reasons and what those deals cost.
  • You've built this before (Deals): for each new lead, the closest past deals by client and words in the title, with their real hours, cost, value and outcome. You can also search on the Deals page.

Seat review

Under the seat optimiser on the Team page, the Seat review gives every paid seat a verdict, for execs and the workspace owner: keep, move up, move down or cancel. Team seats are only kept or cancelled. Open a row for the sentence behind it, what would change it, and the evidence. Every number is computed by Cardinal from the usage Claude Code reports; no model writes any of it.

  • Confidence: high once a seat is 30 days old and the signal is clear; medium when a figure is within 25% of the line that decides it; low when the seat is under 30 days old.
  • Use against the plan: API-equivalent use for the last 30 days and the 30 before, how many times the seat price that is, and whether it is rising, steady (within 15%) or falling.
  • Limit hits: days with a rate limit in the last 30 and 60 days, and blocked time: from each rate limit to the person's next answer from Claude, at most 5 hours, with overlaps counted once. Blocked time is an estimate.
  • Cost of being blocked: blocked hours times the person's cost per hour, against what the bigger plan costs more a month. Payback is the blocked hours a month that pay for the upgrade. It needs a cost per hour on the person and a dollar rate; it is an estimate.
  • Workload: the busiest day's use, the share of use on Opus (when Claude Code reports the model) and the days active.
  • Downgrade safety: a seat only moves down when its busiest day stayed under the smaller plan's price ÷ 30 × 5. One heavy day keeps it where it is.
  • No usage data: a seat whose Claude Code reports sessions but no requests gets no verdict. Check its Claude Code sends telemetry.

Anthropic doesn't publish exact allowances, so headroom is inferred from limit hits.

Later, done and dismiss

  • Later hides an item for 3 days.
  • Dismiss asks why. "Not relevant", "Already done" and "The facts are wrong" keep that subject quiet for 14 days.
  • A note you leave when dismissing becomes a rule, and a draft you edit before marking it sent becomes an example of your wording. Both are used the next time Claude words an item for you. See and delete them on your Me page, under "What your agents have learned". Nothing is used to train any model.
  • When the signal goes away (the deal moved, the invoice was paid), the item closes by itself.

Ask Cardinal

Asking now happens in Cardinal AI: press "Ask Cardinal AI" on Today, or open it from the header. It answers from your own numbers, and it can't see anything you can't. It never changes anything yet. See "Cardinal AI".