How a spend works

1. 01

   The assistant reads the price

   Before a scan, a campaign, a run or a draft rewrite, the assistant calls an estimate tool and gets the credit cost from the server. Prices are the same as in the dashboard; get_credit_pricing lists them.

2. 02

   It asks you

   The assistant tells you the number and waits. It is instructed never to decide a spend on its own.

3. 03

   Your yes travels with the call

   In Claude Code and other clients that support it, the server asks you directly with a confirmation dialog, and the assistant cannot answer it for you. Elsewhere the assistant passes the approved amount as approvedCredits. The server compares it with its own live estimate and refuses the call when the approval is missing or lower than the estimate.

4. 04

   The charge is marked

   Every credit charge started from an assistant shows on your usage page with a via MCP label, so you can always see what the assistant spent.

Limits on every connection

• A request limit per minute and a credit cap per day apply to each connected app. The current numbers are on the MCP page and in get_account_usage.
• The account's own limits still apply: plan credits, the weekly credit cap on each campaign, and the mailbox daily send cap.
• When a limit is hit the tool returns a clear error and the assistant tells you when to try again.

Sending

• A campaign created from chat holds every draft for review. The assistant can read, edit, include and exclude drafts, but send_drafts only queues what you confirmed, and it names the count and the scope first.
• reply_to_thread sends one reply in one conversation, with the text shown to you first.
• Automatic sending on a campaign can be turned on from chat, but only when you ask for it in plain words, and the assistant is told to say so in its reply.