Only people on this list can sign in. Groups decide which tools each person gets (edit them in Access); grey ones come from Google. SQL: the assistant may run read-only SQL for them when no tool fits. Brain: they may confirm SQL answers into the brain (you merge every proposal). Daily limit: dollars a day, empty = the default. Only you can change this list.
Questions no tool could answer, with what was missing and what was tried. Each one is a candidate for a new tool or brain note.
Answers confirmed by people with SQL access, drafted into the brain as a branch of the vault repo: a rule-note section and, when the question will come back, a saved-query tool. Merge it there and it is live within 10 minutes; nothing changes until you do.
What the review agents found: bad and wrong ratings and unanswered questions re-checked against live data, and Customer Care cases the assistant got wrong, grouped by issue. Answer an item to settle it, then propose the answer to the brain; dismiss what needs nothing.
Every run_sql (SAP), run_redshift (product data) and run_soql (Salesforce), from the web app and claude.ai. Each one is a question no tool answered: the share should go down. A query that repeats with other values is a tool to build.
Repeated this month: tools to build
By person this month
Latest runs
Estimated from list prices for every model call (answers, chat titles, the model choice, proposals, review agents). The Anthropic invoice stays the truth. ⚠ = this month above 3× the typical person (and at least $5); you also get an alert.
From claude.ai
Tool calls through the B1 Brain connector in claude.ai. No $ here: those chats run on each person's own Claude seat. ⚠ = more than 3× the typical person (and at least 50 calls). Revoke signs them out of the connector; removing them from People blocks them at once.
Each group gets the tool sets you tick, plus any extra tools. Someone in several groups gets what each allows; Everyone else covers people in no group. You always keep every tool. Changes apply within seconds.
Advanced (JSON)
The same policy as text, for deny lists, signin groups and anything the cards don't show.