Ask Claude About Your Own Money: The LumaLedger MCP Connector

The assistant inside LumaLedger answers questions from your ledger: what you spent, what you owe, what is due, what is left. Since this week the same answers are available from outside the app. LumaLedger now runs an MCP server, so Claude Desktop and Claude Code can ask your ledger the same questions in the middle of whatever else you are doing, with a token you mint in Settings and can revoke in one click.

This post is the launch note. It covers what shipped, what Claude can and cannot do with it, how the token is designed so that handing it out is not reckless, and the five-minute setup.

What shipped

MCP, the Model Context Protocol, is an open standard for connecting AI applications to external systems, and its own documentation compares it to a USB-C port: one plug that works across clients.^[1] In MCP terms a server offers tools that a client such as Claude can call. LumaLedger’s server offers exactly one set of tools: the forty read tools the in-app assistant already uses, with the same names, descriptions and inputs. There is no second implementation. If the assistant can look something up, the connector can too, and when a new read tool ships for the assistant it appears over MCP the same day.

The server speaks MCP over HTTP with JSON responses, accepts batched requests, and authenticates with a personal access token. Claude Code adds it with one command; Claude Desktop adds it as a custom connector, which Anthropic makes available on its Free, Pro, Max, Team and Enterprise plans, with free accounts limited to one custom connector.^[2]

What Claude can ask

The tools map onto the app’s pages. Net worth and its breakdown by asset and liability class. The income statement and cash flow for any window. Budgets, project budgets, savings goals and the debt payoff plan. Recurring charges, the bill forecast for the next quarter or year, the bill calendar and the cancel guide for a merchant. Holdings and investment activity, loans, properties and vehicles. Receipts and paystubs, line by line. Counterparties, rules, scheduled entries, and a transaction search that takes the same filters the list page does. The monthly report and the annual review.

LumaLedger

Assistant

What did restaurants cost me this quarter, and is that up on last quarter?

Income statementTransactions

Restaurants came to $1,412 this quarter across 47 transactions, up 18% on last quarter’s $1,196. Two-thirds of the increase is one merchant, which appeared nine times in August.

Why is my net worth trending down?Where did my money go last month?

The interesting questions are the ones that cross tools, because a client can call several in a row. Which subscriptions renew before payday, and does the checking balance cover them? What did the last three receipts from the hardware store contain, and which project budget were they filed under? Is the savings goal still on pace if this month’s overspend repeats? Each of those is a handful of tool calls that Claude sequences itself, and each answer cites the tool it came from.

What it will not do, on purpose

Nothing over MCP changes the ledger. The in-app assistant can recategorize a transaction, set a budget, name a Venmo payment or mark a transfer, and every one of those goes through a propose-and-approve card: the assistant proposes, you see exactly what would change, you approve. A connector has no way to show you that card, so the actions are not offered to it. This is a design decision rather than a gap, and it will not change until MCP clients can render an approval step we trust.

Two more limits are worth knowing. A token belongs to one ledger: if you are a member of a household ledger and a business ledger, you mint one token per ledger, and no header can move a token from one to the other. And a token never outranks you. It carries your role in that ledger, resolved on every request, so a viewer’s token reads and a member who leaves a ledger has tokens that stop with them.

How the token works

A personal access token is minted in Settings under API & MCP. You name it after what will use it, pick a scope, and see it once. Only a hash of the token is stored; the list shows its first eight characters so you can tell tokens apart, and never the token itself. Revoking it takes one click, and anything using it stops with the next request.

The last row is the one people forget to check when they hand a token to anything. LumaLedger reaps tokens with the account that made them and with the ledger they were made for, so there is no such thing as an orphaned token still reading a deleted user’s data. Deletion here is meant to remove everything, and tokens are part of everything.

Set it up in five minutes

1. 1
   Mint a read token

   In LumaLedger, open Settings and find API & MCP. Choose New token, name it for the client (Claude Desktop, say), leave the scope on Read, and copy the token. It is shown once.
2. 2
   Claude Code: one command

   Claude Code adds a remote HTTP server with a single command that takes the server URL and an authorization header.^[3] The token screen shows the exact line with your URL filled in; paste the token where the placeholder is, and the tools are available in your next session.
3. 3
   Claude Desktop: a custom connector

   On an individual plan, go to Customize, then Connectors, choose the plus button and Add custom connector, and enter the MCP server URL from the token screen.^[2] Add the token as a bearer authorization header where the connector asks for it. Team and Enterprise owners add it under Organization settings instead.^[2]
4. 4
   Ask something only your ledger knows

   A good first question is one with a number you can check in the app: your net worth, or what groceries cost last month. If the two agree, the connector is wired. If Claude answers without calling a tool, the token was not attached.

The same API, for scripts and spreadsheets

There is no separate developer API. The token is a bearer token for the routes the app calls, so anything that can send an HTTP request can read your ledger the way the app does. A spreadsheet that pulls the balance sheet on Monday morning, a script that checks whether a bill posted, a dashboard on a wall: each is one request with one header. Editors and above can mint a write-scope token for a script that needs to post, and it can do exactly what that member can do in that ledger and nothing more. A token cannot mint a token, so a leaked script credential cannot quietly make itself permanent.

The MCP tools are a curated view of that API, slimmed to what a model needs rather than what a screen needs, with ids named for what they are. If a question needs something the tools do not expose, the raw API is a step away with the same token.

Why a ledger is the right thing to hand an AI

Most finance apps hold a list of transactions with categories on them, and an assistant reading that list is estimating. LumaLedger holds a double-entry ledger: every transaction is a balanced journal entry, every balance is the sum of its postings, and the reports are derived from the same entries rather than kept as separate totals. That is what makes the connector worth using. When Claude reports your net worth, it is reading a ledger that ties out, and the number will match the app, the balance sheet and the bank to the cent, because they are all the same postings.

We wrote about why that structure matters before there was an assistant to benefit from it, in double-entry for personal finance, and about the difference between an assistant that reasons over your real records and one that reasons over what you paste into a chat, in what an AI assistant can be trusted with. The connector is where those two arguments meet: the records are sound, and now they are reachable from the tool you already talk to. If you want to know how the app keeps them sound, the reconciliation that runs behind every import is explained in bank reconciliation.