Three steps, then every screen.

1. Connect a source

A read-only key or a sign-in on the vendor's page. The key is tested on seven days before it is stored; the first sync reads ninety and the scan arrives by mail. The screen →

2. Give every dollar an owner

Unallocated is a named number. Attribution suggests rules from the evidence, previews the coverage each reaches, and publishes; every screen recomputes. The screen →

3. Realize a saving

A finding carries its evidence and the change to make. Approve it; the expected saving lands on the ledger, and the verified saving follows after seven days on the same baseline. The screen →

Every screen

Overview

The estate on one screen: the total, coverage, what is realizable now, the governed share, the day-by-day chart with the spike ringed, and the three findings worth most.

Connect

Every source, connected or not: what it reads, what it needs, and what the last sync found. A key is tested on seven days before it is stored, encrypted.

Estate

What is running and for what: every provider, model, workflow and agent, with its share, its owner or its gap.

Spend

The ledger by any dimension — provider, model, workflow, workload, team, agent, metering, status, day — with the change against the window before, both windows named.

Attribution

Where unallocated spend gets an owner: suggested rules from the evidence, a preview of the coverage each reaches, and every rule's history.

Outcomes

Spend against what it produced: each workflow with its outcome, joined by trace or by workflow, and its evidence class.

Unit economics

Cost per successful outcome, per workflow, with retry waste and failures kept apart from the successes.

Opportunities

Every finding the detectors raised, with its evidence, the change to make, how it is verified, and what was decided. Expected savings by kind, and every finding by state.

Scenarios

Change one assumption and watch the forecast move: the same runs re-priced on another model, a cache TTL, a retry cap, a seat count.

Ledger

Every decision on a finding: the saving expected on approval, the saving measured after the change, against the same baseline. Append-only.

Agents

Every agent with its business owner, its monthly envelope and its mode. A gap is named, never filled.

Policies

Templates with defaults from your own ledger, backtested on the window before a mode changes, and a ladder from observe to enforce.

Approvals

The held batch: runs waiting once a policy in Approve has breached, who decides, and the record of each decision.

Alerts

What the seven watches raised in the window, each with its severity, its cause and where it is decided; addressed or ignored with who and when; whether it was sent.

AI P&L

The report finance keeps: spend by team and workflow against the month before, coverage, realized savings, and the export — CSV and JSON from the same builder.

Ask Valistry

A question in plain words, answered with the figure, the plan that produced it and the rows behind it. The planner reads; the engine computes; a model narrates and never does arithmetic.

Reference

The words

Every term on a screen, in a mail or in an answer, defined the way a test could hold it.

The record

Every field one usage record holds, and the six things it never holds.

Connectors

For each source: what it reads, the access it needs, what it attributes on its own, what it cannot answer.

The read API

The same figures the screens show, as JSON, with the key or the sign-in the editor uses.

llms.txt

All of this as one text file, for an agent that reads before it signs in.