Kelpie reads your team's mail and calendar and scores every account from what the customer actually did. The working is printed next to the number. Your CSMs get the answers inside Claude or ChatGPT. You get a book you can defend to the board.
Reads email and calendar only · Never stores a message body · One database per customer
| Account | Health | Silence | Ball | Owner |
|---|---|---|---|---|
| Harbour Optics | at risk · 40 | 43 business days | us | Priya |
| Ridgeline Outdoors | at risk · 40 | 29 business days | us | Tom |
| Wattle & Co | attention · 20 | 11 business days | them | Priya |
| Pelican Press | healthy · 0 | 2 business days | them | Mei |
The number, the reason, and the message it came from.
Every CS team has the spreadsheet. Green, amber, red, updated when someone remembers. It reads healthy right up to the renewal call, where the customer mentions they decided months ago.
The first time Kelpie scored a real book it disagreed with the spreadsheet on 20 of 62 accounts. One account sat at “20, fine” after 29 business days of silence. Another read “healthy” while the customer had not written back in 19 business days, because the CSM was sending weekly check-ins and the spreadsheet counted those as contact.
Nobody was lying. The number just had no evidence behind it, so there was nothing to argue with and nothing to act on.
Silence is counted in business days since the customer last wrote, and only while your side sent the last message. Chasing an account does not make it look healthier. An out-of-office is not a reply. A usage crash is a fixed number of points, whether or not the CSM likes the customer.
A CSM who knows something the mailbox cannot see records it with a quote and a message ID, and the score moves by a fixed, published amount. It expires on a review date. Nothing on the page can be nudged green.
What that means on a Monday. The conversation goes from “why is this amber” to “what are we doing about the 43 days”.
Kelpie is a remote MCP server. A CSM connects it once to Claude, ChatGPT or any client that speaks the protocol, signs in, and from then on asks:
/prepBefore a call. Health with the working, the last five meetings distilled, every promise still open and the reply that may have settled it.
/todayAt the start of the day. Meetings, commitments due, a notice deadline landing today.
/mondayFor the week. The book, worst first.
/housekeepingOnce a week. The questions only a human can answer, one at a time, fifteen minutes.
A 7am email carries the same lists for people who would rather not ask. Managers get a dashboard: every number, the evidence behind it, and the questions nobody on the team could answer. Judgement goes in through the agent, where the person entering it can be asked a follow-up.
Every recorded call and every matched customer email is read once and reduced to single observations. A decision. A risk. A contact change. A request: something the customer asked for that we do not do. A commitment, with a due date and a name against it:
Roy will send the export by Friday.
That fact is never edited. If Roy sends it, a new fact says so and the old one stays as the record of what was promised. If he does not, it shows as overdue and waits for a person to say what happened. Kelpie never decides a promise was kept.
Measured, on one book. A year of customer email became about 1,400 facts for about sixteen US dollars, and surfaced 32 overdue commitments that had only ever existed in email.
A renewal date, a next step, the reason an account is at risk: those are somebody's words, recorded with their name against them, and Kelpie asks rather than guesses. A guessed renewal date is worse than none, because everything downstream treats it as fact.
It does not read chat, your CRM or your support desk. Where your customers talk to you somewhere Kelpie cannot see, you say so at onboarding and it says so on the page rather than reading silence as trouble.
It does not send anything to your customers. Ever.
Each night Kelpie ingests and scores, then works out what it cannot answer on its own: an account with no contract on file, a lifecycle stage the mailbox disagrees with, a promise that looks discharged. On Monday it asks the account owner those questions inside their assistant, one at a time, and records every answer under their name. “I don't know” is an answer. Asked twice, the question goes to their manager.
The first real session: four questions, four answered, the first three in 95 seconds. Two sessions later the open questions on that book had gone from 45 to 26.
Message bodies are fetched at the moment they are read, turned into observations, and dropped. There is no body column. The storage layer has nowhere to put one.
A mechanical check refuses any fact carrying a currency symbol or a verbatim quote. Across every mail fact on the first book, the count of either is zero.
Every customer gets their own database, not a customer column in a shared table. Row-level security stays on underneath as a second wall. 406 tests walk the catalogue on every build.
The rehearsal brought back 59 accounts, 15,355 records and 6,568 facts, matched row for row. Point-in-time recovery is on, so a bad write at 2pm is recoverable to 1:59pm.
Google Cloud, Australia. Sign-in through WorkOS. No passwords stored.
A seat is a connected mailbox, not a login. Managers and admins who read the dashboard are free, so nobody rations the one surface that should be open. Single sign-on through your own Okta or Entra is AUD 250 a month on top. Pay annually and two months are free.
The walkthrough is thirty minutes on your own book, after a short setup: connect one mailbox, upload your domain list, wait one night. You see your scores with their working the next morning. Nothing is written to your systems.
Book a walkthroughOr connect an agent now and ask it about the demo book:
claude mcp add --transport http kelpie https://mcp.kelpie.pro/mcp