Let your own agent read the ledger
In the early eighties, at British Steel, a new Apple arrived in the industrial engineering department and nobody knew what to do with it, me included. It had VisiCalc on it, and I could see straight away that it was going to put the A3 analysis paper and the Tipp-Ex out of work. I’ve been looking for the next pot of Tipp-Ex ever since, and in a bank it isn’t hard to find. It’s the month-end reconciliation. Someone exports the subledger, someone else exports the general ledger, and a capable person spends the first days of the month ticking one against the other in a spreadsheet. Every bank has that person. They’re usually very good at it, and it’s usually a waste of them.
What MCP is
The Model Context Protocol is a standard way for an AI agent to talk to a system it has never met. Anthropic published it in late 2024, and it has since become the thing most agent tools speak, Claude and Claude Code among them. A system that speaks MCP tells the agent what it can do, as a list of tools with a description of each, and the agent decides which to call.
That sounds dry, so here’s why it matters. For as long as I’ve been in this business, the answer to “can your system talk to ours?” has been a connector project: a statement of work, an integration team on each side, and a few months. With MCP, nobody writes an integration for each pair of agent and system. You give the agent an address and a credential, and it works out the rest from the descriptions.
The door
LedgerTM’s MCP door is that address, one per organisation. Your agent connects, signs in, and asks what it may read. It gets a tool for every list its permissions allow, such as journal entries, account balances or transactions, along with a search to find the right one, a tool that groups and totals rows the way our own assistant does, and every public guide article as text. If your organisation has the reporting tables, the agent can also write its own SQL against them, with a dictionary of every table and column to steer it.
A person can connect through their usual sign-in, in whatever agent tool they like, and the agent reads with that person’s permissions. An agent that runs on its own, at three in the morning with nobody watching, gets its own credential from your identity provider and its own group, and I’d give it the group a clerk doing the same job would have.
Either way, it only reads. No tool posts, approves or edits anything, and an agent can never approve even if its group would allow it. The door uses no model of its own, so nothing goes to our model vendor: your agent chose its own model, and the rows go wherever your organisation sends them. Every call is kept under the caller’s name, on the same review screen as the assistant’s conversations, so “what did the reconciliation agent read on the 5th?” is a question with an answer.
The GL Reconciler
This is the part I’m most pleased about. Anthropic publishes a set of finance agents, Claude for Financial Services, which you can install as plugins in Claude Cowork or deploy behind your own workflow through their Managed Agents API. One of them is the GL Reconciler. You give it a date, and it pulls balances from the general ledger and the subledger, finds the breaks, traces each one back to the transactions behind it, has a second agent check its work, and writes an exception report for a controller to sign off. It doesn’t post a thing, and adjustments stay with a person, which happens to be the rule we hold our own assistant to.
What it asks for is two read-only MCP servers, one for the general ledger and one for the subledger. LedgerTM is the subledger, and the door is a read-only MCP server. So the reconciliation that used to be an export, an import and a week of someone’s month becomes an agent your controller starts with a date, reading your general ledger on one side and us on the other, and nobody wrote an integration to make it happen. It used to be a connector project. Now it’s an address and a client id.
I won’t pretend it’s turnkey on the first morning. Anthropic’s agents are starting points, written with fund accounting in mind, and you’ll want to tune the prompt and the thresholds to the shape of your book. Your general ledger needs an MCP server of its own too, from its vendor or your own team. But the part that used to take months is now a standard both sides already speak, and the person with the spreadsheet gets their first week of the month back.
Switching it on
The door is closed until your organisation switches on Platform › MCP Door, and a closed door answers as if it didn’t exist. Using it is a permission granted per group, and an agent can reach exactly what its group can. It’s built, tested and goes out in the next release. The guide article on the MCP door has the connection steps.
Written by Jeremy Rich.