Your AI Assistant Can Now Do Credit Card Math Correctly

Ask an AI assistant which card to use at Costco and it will usually tell you to use your grocery card. That is wrong, and the reason it is wrong is instructive.

Credit card arithmetic is simple. The rules wrapped around it are where answers go wrong. Costco codes as a wholesale club, so grocery bonuses skip it. The 5% category usually carries a $1,500 quarterly cap that needs activating first. "8x on travel" often means 8x through the issuer's portal and 4x booking direct. Models handle the multiplication fine and the rules badly, because the rules are the part nobody remembers correctly.

CardSavvy now runs an MCP server. Your assistant can call our optimizer directly instead of recalling anything.

Or just use the optimizer yourself →

What this is

Model Context Protocol is a standard way for an AI assistant to call outside tools. Connecting CardSavvy gives your assistant eleven functions it can run: look up a card's real terms, check how a merchant codes, score a single purchase across your wallet, or solve a full year of spending with the linear program that powers the optimizer.

The model stays entirely in your client. There is no field anywhere on card-savvy.com to paste an sk-ant- or sk-proj- key into, because we never needed one. Your assistant decides that classify_merchant and best_card_for_purchase are the right calls; our server runs them and hands back numbers.

We log the tool name, a SHA-256 hash of the inputs, a cache flag, and the elapsed milliseconds. Not the arguments. Your five card names and your $9,600 of annual grocery spend go through the calculation and are never written down.

Every answer carries an as_of date, a source_url pointing at the issuer's own page, and the line about this being educational rather than advice. That last one matters more here than on a web page, because you will never see our footer.

Getting started

The catalog, merchant-coding, and single-purchase tools need no account. One command:

claude mcp add --transport http cardsavvy https://mcp.card-savvy.com/mcp/

For Claude Desktop, add it to your MCP settings instead:

{
  "mcpServers": {
    "cardsavvy": {
      "type": "http",
      "url": "https://mcp.card-savvy.com/mcp/"
    }
  }
}

Then ask it something it would previously have guessed at:

I have a Sapphire Reserve and an Amex Gold. I'm buying $200 of groceries at Costco. Which card?

It will check how Costco codes, score both cards, and tell you the Amex earns more here while warning you that Costco only takes Visa in the warehouse. That second part is the kind of thing a model answering from memory reliably forgets.

For the wallet optimizer, create a free token at your profile under API Access. It requires a verified email address, and the token is shown once. Add it as a header:

{
  "mcpServers": {
    "cardsavvy": {
      "type": "http",
      "url": "https://mcp.card-savvy.com/mcp/",
      "headers": { "Authorization": "Bearer cs_mcp_..." }
    }
  }
}

Now you can ask the questions that need real math:

Here are my five cards and roughly what I spend by category. Which card should I use for what, and is the Reserve still worth its fee?

Without a token you get 120 tool calls an hour. With one you also get 60 optimizer solves an hour and 300 a day, which is far more than a person asking questions will ever use and far less than a runaway script needs. Full documentation and the whole tool list is at card-savvy.com/mcp.

What it will not do

Every tool is read-only. Nothing here opens a card, closes one, changes an account, or submits an application. It reads data and does arithmetic.

It will also not optimize around a problem that outranks optimizing. If you tell it you carry a balance month to month, the response says so before anything else. Interest at typical card APRs costs several times what any rewards rate earns, and a tool that answered "here is your optimal allocation" to someone paying 24% APR would be technically responsive and practically useless.

Treat the output as information rather than as advice. It is arithmetic on numbers you can check, each carrying a link to where it came from.

Why this and not a chatbot

A chat box on card-savvy.com where you paste an API key would have been easier to build. Three reasons we did not.

Our database has no column for an API key, and never has. Every secret we hold is a hash: passwords, refresh tokens, and now the cs_mcp_ tokens, which are SHA-256 and shown exactly once. Adding a table of other people's live credentials to a site with no revenue would be the single worst trade available to us.

Running the model ourselves costs money per conversation. A $0 product that pays per answer either starts charging or gets quietly worse, and you would be paying twice, because you already pay for Claude or ChatGPT.

The third reason decided it. If we generated the prose, our arithmetic and a language model's output would arrive tangled together with no way to tell which part to trust. Here the assistant orchestrates and never computes. When it says the Reserve nets $957 a year on your spending, that number came from a linear program you can call yourself and check.

Quick decision guide

You use Claude, ChatGPT, or another MCP client and hold more than two cards: connect it. The single-purchase tool alone will change what you reach for at the register, and it costs nothing to try.

You want the whole-wallet answer: create the free token. The optimizer solves category-by-card allocation net of annual fees, which is not something you can eyeball.

You are building a product: the API terms permit commercial use and there is no fee. Tell us what you are building at support@card-savvy.com, mostly so we can warn you before anything changes.

You would rather not connect anything: the optimizer has always worked in a browser and still does. This changes nothing about it.

Frequently Asked Questions

What is an MCP server?

Model Context Protocol is a standard way for AI assistants to call outside tools. Connecting CardSavvy gives your assistant a set of functions it can run, like looking up a card or solving a wallet, instead of answering from whatever it remembers about credit cards.

Does CardSavvy see my AI conversations?

No. Your conversation goes to whichever model provider you use, under their privacy policy. What reaches us is one tool call: card names, a dollar amount, a category. Asking "should I keep the Reserve" sends us the string "Chase Sapphire Reserve" and your category totals, and nothing else from the conversation.

Is it free?

Yes. Card lookups, merchant coding, and single-purchase scoring need no account at all. The wallet optimizer needs a free token you create from your profile, which exists so we can apply a per-integration rate limit and know who is building on this. There is no paid tier.

What does CardSavvy log?

The tool name, a one-way hash of the inputs, whether the response came from cache, and how long it took. We do not log tool arguments. Your wallet and spending pass through the calculation and are not written down.

How current is the card data?

Every response carries an as_of verification date and a link to the issuer page it came from. Twenty-one of our 137 cards currently have a verification date; the rest report as unverified rather than pretending to be current, and your assistant can tell you which is which.

Can I use this in a product I am building?

Yes, the API terms permit commercial use and there is no fee today. If you build something on it we would genuinely like to hear about it at support@card-savvy.com, partly so we can warn you before anything changes.

---

CardSavvy earns no advertising or affiliate revenue today. The recommend_cards tool strips the apply_url and affiliate_partner fields before answering and returns the issuer's own page instead, so a card ranks above another only because the arithmetic put it there. If that ever changes we will say so, and it still will not change the math. We wrote down why before there was anything to be tempted by.