This was on HN a couple weeks ago. The idea is to let an LLM classify your bank statements into categories, but tally has an interesting approach. It's not just a prompt, and its not an app calling the LLM internally; instead, the your agent is supposed to call the app.

tally basically wants the agent to build a set of rules to match against your transactions, so it keeps telling the agent "call this subcommand to get some sample transactions that match no rules yet", the agents will generate rules that seems sensible, then invoke tally again to validate their work; tally itself merely implements the rule engine.

The agent still needs to be able to build a view of your categories primarily, and then your existing rules as well I suppose - because in some cases the right solution might be to update an existing one; I could see this becoming an issue as your ruleset grows. tally gives it the backpressure it needs.

I quite like the idea of externalizing the control loop to the user's existing agent of choice. It makes the entire thing feel instantly incredibly powerful compared to something that uses an LLM internally but gives you a fixed interface to work with. For example, you are able to tell Claude Code to make adjustments like "can we split this category in two" and it would work flawlessly. It also makes tools composible!

There is also a command (tally workflow) prints instructions to explain to agents how the process works. This reminds me of the beans prime subcommand. A pattern in the making!