AI & agents

A chat-focused shell with suggested prompts grounded in the household's own accounts, a streaming answer whose tool steps run in view, and the rows it read kept in the evidence pane beside the thread.

A spending answer where every number links back to the transactions behind it — citations interleaved into the prose, and the sources themselves in the evidence pane.

AI suggestions with confidence scores and one-tap accept or reject — never auto-applied, and the day's decisions summed where the row stops.

A proposed money move run on the focus frame: the agent's plan shown before it runs, the reason stated, full row details, and an explicit approve or deny.

Queries, tokens, cost, and saved hours against the plan's allowance, with a by-feature split and the months behind it.

One finished run taken apart: every step with its result and duration, every tool call with what matched and what changed, and the four rows it moved.

The explicit list of what the assistant reads and what it never reads, each source toggleable, with the data handling stated in the same breath.

A recurring briefing the assistant runs on schedule — next run, last runs with their statuses, and an honest pause rather than a delete.

A thread with a statement held alongside: per-file provenance for what was read, what it said, and what was skipped, before the answer.

The assistant's own ledger — every query, tool run, and action with its approval, its cost, and its duration, filterable like any other book.

The same auditable thread in one thumb-width column — the answer, the rows behind it folded away, and a composer that never covers what it just said.

The four honest answers: outside what it reads, too little confidence to say, needs a person to approve, and the allowance spent — each with the way forward.