A Free AI Agent Just Learned to Do Small Business Bookkeeping

Your bookkeeper never needed the cloud. Neither does your AI.

The shoebox on the desk at 11pm

Priya runs a six-person catering company outside Manchester. Every month ends the same way. Wednesday night, the shop closed, the shoebox of receipts comes out. Fuel. Ingredients. A last-minute van rental for the wedding job in Cheshire. She types each one into a spreadsheet, guesses at a category, and emails the mess to her accountant on Thursday, three days later than she promised.

This is not a Priya problem. It is the default state of small business bookkeeping everywhere, from a landscaping crew in Ohio to a boutique in Lagos to a five-van plumbing outfit in Leeds. The receipts pile up because reconciling them by hand is boring, and boring work gets pushed to the end of the day, then the end of the week, then the accountant's inbox at midnight.

You are not bad at bookkeeping. You are the bottleneck.

Here is the quiet admission most owners never say out loud: the numbers were never the hard part. Typing them in was. Every hour spent matching a receipt to a bank line is an hour not spent on the phone with a client, not out quoting the next job, not asleep.

For years the honest answer was "hire someone" or "buy software and do it yourself anyway." Now there is a third option, and it showed up three days ago with almost no warning.

What is OpenWorker, and can it actually do your books?

OpenWorker is a free, open-source AI agent that Andrew Ng and Rohit Prasad released on 23 July 2026 under an MIT licence. It is not a chatbot. Ng's own words on the launch: it "doesn't just chat with you, but delivers finished work." You give it a goal in plain English, such as "read the receipts in my Downloads folder and this month's bank export, categorise every transaction, and draft a profit-and-loss summary for my accountant." OpenWorker plans the steps, reads the files on your own computer, does the categorising, and hands you a finished document.

It runs as a desktop app on your Mac today, with Windows support in beta. Under the hood it is built on aisuite, the open-source library Ng and Prasad have maintained for over a year for calling any large language model through one interface. That is why OpenWorker is not locked to one AI provider. Point it at OpenAI's GPT-5.6, Anthropic's Claude Fable 5 or Opus 4.8, Google's Gemini 3.6, an open-weight model such as Kimi, GLM, DeepSeek, Qwen, or Mistral, or run it fully local through Ollama with no API key and no per-token bill at all.

Isn't it risky to hand an AI agent your bank statements?

This is the objection every owner raises first, and it is the right question to ask. The honest answer is that OpenWorker was built around exactly this worry.

The agent runs a local server on your own machine. Your files, your model API keys, and your conversation history sit in that local server's own secret store, not on a vendor's cloud. The only traffic that leaves your device is the specific instruction sent to whichever AI model you chose, and the login handshake for any app you connect. If you run a fully local model through Ollama, nothing leaves your laptop at all, not even that.

OpenWorker also uses four permission levels: read, write local files, run a command, and send something external like a Slack message or an email. The last two, running a command and sending something outside your machine, always pause and wait for your explicit approval before they happen. It will draft the email to your accountant. It will not send it until you say so. That single design choice is the difference between an assistant and a liability.

How OpenWorker actually closes your books

This is the same four-step pattern Ng highlighted at launch, applied to a monthly reconciliation instead of a customer brief:

1. Install and choose your model. Download OpenWorker from GitHub, then decide how you want to pay for the intelligence: bring an API key for Claude, GPT, or Gemini and pay per task, or install Ollama and run a model fully free on your own hardware.
2. Point it at your financial mess. Give it read access to the folder where your receipts, invoices, and bank export CSV live. Nothing uploads anywhere until you connect a tool or approve a send.
3. State the outcome, not the steps. Tell it what you want finished, not how to do it: "categorise every transaction in this month's export against these expense categories, flag anything over 200 pounds for my review, and draft a one-page profit and loss summary."
4. Watch it plan and work. OpenWorker breaks the goal into steps, reads the files, and works through the categorisation, the way a competent junior bookkeeper would on their first day, minus the training period.
5. Approve before anything leaves your machine. When the report is ready to email to your accountant or post into your Slack channel, OpenWorker stops and asks. You read it once, then say go.
6. Schedule it and forget about the shoebox. Set it to run the same routine on the first of every month. Scheduled jobs keep a full transcript, so you can always see exactly what it did and why.

How does this compare to what you're already using?

Google's own AI Overview on this exact question, checked the week OpenWorker shipped, says fully autonomous free bookkeeping agents are still rare, most zero-cost tools lean on basic automation rather than a real agent that plans and acts. Here is where OpenWorker actually sits next to the alternatives it names:

What this looked like for Priya's catering business

Priya installed OpenWorker on the Wednesday it hit her feed. She pointed it at Ollama running a small local model, since her books hold nothing she wanted touching a cloud API by default, and gave it read access to her Downloads folder and a CSV export from her bank. Her first instruction: "match every transaction to a category, flag the van rental and anything over 150 pounds, and draft a one-page summary for my accountant."

Twenty minutes later she had a draft. Twelve transactions were flagged for her to glance at, not four hours of typing. She read it once, corrected one miscategorised fuel charge, and approved the email. Thursday's deadline, met on Wednesday night, for the first time in eight months.

What you have after this

A week in, you have stopped opening a spreadsheet at all. A month in, your accountant is getting a cleaner report earlier than they have ever received one from you, and asking what changed. Six months in, the shoebox is gone, because you photograph receipts straight into the folder OpenWorker watches, and the monthly close runs itself while you sleep.

Same business. Same accountant. A completely different Wednesday night.