Latest experiments with AI and Beancount?

[Joshua Cabrera]

Hi all,

Does anyone care to share any tinkering they've done with AI tools recently?

I found Tammo's post sharing a project for a bean-query MCP server which is quite interesting.

I've been using Beancount for just over a year, but as a layman, I've struggled to automate my workflow.

I was thinking of installing a local model to explore benefits, which is how I found Tammo's post.

Also, would anyone share thoughts on how well or poorly this might go?

• share my existing ledger for inference,
• share an account statement as the raw transaction input I wish to record, and
• prompt the model to output the proposed Beancount transactions that I can review and then save to my ledger.

It seems like this would work well by iterating on the prompt, but I would appreciate any feedback before I potentially lose precious weekend hours on this idea.

Cheers,

Josh

[Jonah Benton]

I have done a little work to do plain text extraction from PDF statements, do a little cleanup, pass text to a local model to structure into data (json), prior to converting to BC transactions. Statements that use different columns, for instance, for debits and credits are kind of hard to deal with reliably with deterministic rules, but in some experiments so far models like 32b sized Qwens have done quite well. With local models would just be careful about how much they are asked to do, and how precisely they are asked to do it.  

--
You received this message because you are subscribed to the Google Groups "Beancount" group.
To unsubscribe from this group and stop receiving emails from it, send an email to beancount+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/beancount/4f44c92f-f7d0-4306-8b86-dd1659257e99n%40googlegroups.com.

[Alen Šiljak]

Interesting topic. I think I raised something similar in the Discussions in the Beancount repo.

I haven't done much for real but was happy to experiment and see what is possible. Admittedly, I don't see much need but I'm trying to use AI agents for what we used to try to solve by coding. Downloading statements, parsing them, even generating entries, now seems easier and more adaptable to changes.

I've used Gemini CLI mainly and it was doing the job. Now I'm looking at some agentic AI frameworks.

Anyway, in summary, I got the model to find the ebanking url of the given bank, open the page, wait for me to log in, then locate the credit card account and how to download the latest statement. It then downloads the PDF and compares the transactions to the ones in Beancount journal. Beancount and Playwright MCPs were enough for this.

Locally, I'd like to try with something fast, like phi4-mini or gemma3n as, apparently, using agents (smaller tasks and focus) achieves greater accuracy than using one model to do everything.
Would be interested in learning more stories. Soon we might just be sharing prompts instead of building plugins and importers.

Cheers