Crew P&L

Reporting, not authority. Nothing on this page approves a spend, releases an escalation, changes a cap, or moves a token. It reads three meters and puts them on one period.

"What did this crew cost last week?" currently means opening Treasury for USDC intents, Billing for operation counts, and the audit trail for connector calls, then reconciling three windows by hand. A crew P&L answers it in one read:

lacrew pnl --crew 0x2222…2222 --period calendar_week

0x2222…2222  2026-W31  (2026-07-27T00:00:00.000Z → 2026-08-03T00:00:00.000Z UTC)
  asOf        2026-07-31T12:00:00.000Z
  Onchain     50.00 USDC spent · 75.00 pending · 200.00 granted
  Inference   $4.5000 · 128 call(s) · 412300 in · 96100 out  (3 unpriced — this is a floor)
  Connectors  46 call(s) · 7 write · 39 read · 1 failed · price unknown
  Budget      hard · ok · $195.50 left of $200.00 (period 2026-07)

What the lines mean

Line	Source of truth	Denominated in
Onchain spent	`AllowanceSpent` / `ActionExecuted` on the audit trail, deduped by transaction	the asset's own base units
Onchain pending	escalations created and not resolved inside the window	the asset's own base units
Onchain granted	`AllowanceStreamed` — what the epoch put into a seat's allowance	the asset's own base units
Marketplace	`MarketplacePurchase` that settled — listed apart from ordinary spend	the asset's own base units
Inference	the F2.28 counters, per metered call	tokens, and best-effort USD
Connectors	`ToolCalled` rows, by connector and route	calls; USD only where a price table says so

Onchain budget ≠ inference budget. They are shown side by side and never added together: one bounds funds leaving the Treasury under a policy stack, the other bounds what the crew's model calls cost you. See Inference cost budgets.

Honesty rules

These are the reasons a figure may look "missing", and each one is deliberate.

price unknown is never $0. A model call no provider priced and no price table covers is counted as a call, its tokens are counted, and the $ figure says how many calls it omits. The same holds for connectors: with no price table configured, connector usage reports call counts and states that the price is unknown.
available: false is not a zero. If inference metering is not configured, the report says model cost is unmeasured rather than showing $0.00.
A partial window says so. Without a database, the orchestrator answers from a bounded in-process ring; the report is flagged incomplete because a window that quietly forgot its oldest rows is a lower bill than the one you will be sent.
Seat rows that do not sum are named. A model call charged to the crew without naming a seat lands in unattributed and is called out in the notes, rather than being spread across seats that did not incur it.
Pending is not spent. An escalation awaiting a decision has moved nothing, and it is listed on its own line.

The period

UTC, always. A workspace timezone would move month boundaries, and a figure that changes with the reader's timezone is not one an accountant can check.

`--period`	Window
`calendar_month` (default)	the UTC month `now` falls in
`calendar_week`	the ISO week (Monday-anchored)
`epoch`	one epoch, from `--epoch-seconds` / `--epoch-anchor`
`--from` / `--to`	exactly those instants; wins over `--period`

Windows are half-open ([from, to)), so two adjacent periods never both claim a row that landed on their boundary. A range longer than 366 days is refused rather than silently truncated.

Pricing connector calls

Nothing ships pre-priced — LaCrew does not know what your workspace pays GitHub or Notion. Supply a table and the routes it covers get a $ figure; the rest stay honestly unpriced:

export LACREW_CONNECTOR_PRICES='{"github.merge_pr":0.01,"notion":0.002}'

The narrower key wins (<connector>.<route> over <connector>), and a negative or unparseable entry is dropped rather than applied.

HTTP

GET /pnl?crewId=0x…&period=calendar_month
GET /pnl?crewId=0x…&agentId=0x…&from=<ISO>&to=<ISO>
GET /pnl?crewId=0x…&period=calendar_week&format=csv

format=csv returns the same aggregate as a flat, one-row-per-line-item export; the usd cell is left empty where nothing could price the row and price_known says which of the two a blank means, so a spreadsheet cannot sum an invented zero.

On the cloud the same report is served per crew and per seat, scoped to the workspace's own org chart:

GET /v1/crews/:id/pnl?period=calendar_month
GET /v1/agents/:id/pnl?from=<ISO>&to=<ISO>

An address that names no node in the caller's org answers 404, not 403 — whether another workspace runs a crew at that address is not information this caller has a claim to.

What it does not do

It is not double-entry accounting and it is not an ERP export. It is what an operator needs to answer "is this desk inside both of its budgets?".
It does not replace Stripe invoices, and it does not bill anything.
It does not stream. Period aggregates carry an asOf stamp and may be seconds behind; polling is enough for a report nobody acts on.
It never widens authority. A P&L that could act would be a second enforcement path for money, and there is exactly one of those — the chain.