AI Agent Bookkeeping: Tracking Money in Plain Language

Most people who give up on tracking their spending do not give up because they stop caring. They give up because entering every transaction is tedious. Open the app, tap the plus button, type the amount, pick a category, pick an account, save. Multiply by several transactions a day and the habit dies within weeks.

Agent-based bookkeeping removes most of that friction. Instead of filling in a form, you tell an AI assistant what happened, in the same words you would use with a friend, and the assistant writes the entry into your ledger. This article explains how that works, what it looks like day to day, and, most importantly, how to do it safely. It is written by the POMODIARY team, who build an open-source tool in this space, but the principles apply to any setup.

What agent-based bookkeeping is

A traditional budgeting app is a database with a form in front of it. An agent-based system keeps the database but replaces the form with a language model that can act on your behalf. The model's job is translation: turn "coffee 4.50 on the Visa" into a structured record with a date, an amount, a category, an account and perhaps a tag.

Three pieces are involved:

The ledger. A database of accounts and transactions. This is the source of truth and should be something you can inspect directly.
The assistant. A chat-based AI such as Claude or ChatGPT, or an agent framework, that you already talk to.
The connection. Some way for the assistant to read from and write to the ledger, with credentials that identify you.

The assistant never "remembers" your finances. It reads them from the ledger when needed and writes changes back. That separation is what makes the approach trustworthy: the ledger can be audited regardless of what the model said in conversation.

How the connection works

There are a few common patterns for connecting an AI assistant to an external system.

Tool protocols such as MCP

The Model Context Protocol (MCP) is an open standard for exposing tools and data to AI assistants. A ledger could run an MCP server offering tools like "add transaction" or "list transactions for a month", and any MCP-capable assistant could call them. The assistant sees a description of each tool and decides when to use it.

Instructions plus an API key

Another approach is to give the assistant a document that describes how to talk to the ledger, together with an API key that authorizes access. This is the approach used by our own tool, Capital: you generate an API key in Settings and copy a generated Skill document into your AI tool. Claude can use it through Projects or a Skill, ChatGPT through Custom Instructions, and OpenClaw through a native Skill. Each user gets an isolated database, and the agent works with it through SQL directly, which means everything it does is a plain, inspectable query rather than a black box.

Which is better?

Neither is universally better. Tool protocols give the ledger tight control over which actions are possible. Direct database access is more flexible and transparent but relies more on careful permissions and review. Whatever the method, the questions to ask are the same: what can the assistant read, what can it change, and how do I see what it did?

What entries look like in practice

The appeal is that you can be terse. Some examples of what you might type and what a well-configured assistant should record:

You say	Recorded as
lunch 30	Expense, 30, Food, default account, today
coffee 4.50 visa	Expense, 4.50, Food, Visa card, today
groceries 86.20 yesterday, tag household	Expense, 86.20, Groceries, default account, yesterday, tag: household
salary 4200 into checking	Income, 4200, Salary, Checking, today
moved 500 from checking to savings	Transfer, 500, Checking to Savings
Netflix renews monthly at 15.49	Recurring subscription, 15.49, monthly

The assistant can also answer questions from the same data:

"How much did I spend on food this month compared with last month?"
"List everything tagged travel in September."
"What subscriptions do I pay for, and what do they add up to per year?"

Questions like these are where an agent shines, because you do not need to learn a report builder. You ask, it queries, it answers. Still, for anything important, open the actual report and confirm.

Safety: the part that matters most

Letting software write to your financial records deserves the same care you would give a new bookkeeper. These are the habits we recommend.

Separate read from write

If your setup allows different permission levels, use them. An assistant that can only read is useful for questions and reports with almost no risk. Grant write access when you want it to record transactions, and be very cautious about anything that can delete or bulk-edit. If a single key grants full access, keep it in one trusted place rather than pasting it into every tool you try.

Confirm before it commits

Ask the assistant to echo back what it is about to record: amount, date, category and account. A one-line confirmation catches most mistakes, such as 30 interpreted as 300 or yesterday's purchase dated today.

Review weekly

Set aside fifteen minutes a week to scan new entries in the ledger itself, not just in the chat. Look for duplicates (the assistant recorded the same purchase twice because you mentioned it twice), wrong categories and wrong accounts. Then compare account balances with your bank.

Treat the API key like a password

Anyone with the key can act on your ledger. Do not share chats that contain it, do not commit it to code repositories, and revoke and regenerate it if you suspect it has leaked.

Be careful with what you paste

Avoid pasting full bank statements, account numbers or card numbers into a chat. The assistant needs amounts, dates and descriptions, not your credentials.

Keep the ledger as the source of truth

The conversation is a convenience; the database is the record. If the assistant says something that disagrees with the ledger, trust the ledger and investigate.

Why open source and self-hosting help

Financial data is among the most sensitive information you have. Open-source bookkeeping tools offer a few concrete benefits:

Inspectability. You or someone you trust can read the code and see exactly what happens to your data.
Portability. If the hosted service changes or disappears, you can run the same software yourself and keep your records.
Control. Self-hosting means your ledger lives on a machine you control, with your own backups.
No lock-in. Data in a standard database can be exported and queried with ordinary tools.

Capital is released under the Apache 2.0 license and can be self-hosted, which is why we chose direct, transparent database access over a hidden service layer. Self-hosting does come with responsibilities: you handle updates, security and backups, so it suits people comfortable running a small server.

Common mistakes

Skipping the review. Natural-language entry makes recording easy, which makes it easy to stop looking. The weekly check is what keeps the data accurate.
Inconsistent wording. If you say "food" one day and "restaurants" the next, you may end up with overlapping categories. Decide on a short category list and stick to it.
Mixing transfers and spending. Moving money between your own accounts is not an expense. Make sure transfers are recorded as transfers.
Expecting advice. A bookkeeping agent records and reports. It is not a financial advisor, and you should not treat its comments on your spending as professional advice.

Getting started

Pick a ledger you can inspect directly, hosted or self-hosted.
Create your main accounts and a short list of categories.
Connect one assistant you use daily, with the narrowest permissions that work.
Record a week of transactions by chat, asking for confirmation each time.
At the end of the week, review everything in the ledger and reconcile with your bank.

If that first week goes smoothly, you have something most budgeting apps never achieve: a habit that survives a busy month.