Skip to main content
Once an assistant such as Claude, ChatGPT or Cursor is connected through the MCP server, you run outbound by asking for it. This guide gives requests that work, says what each one costs, and shows where the assistant should stop and ask you.

Before you start

  • Connect the app. See Connect an app.
  • Give it a daily cap. On the approval page, set the most the app may spend in a day. A cap is the firm limit: a paid call that would pass it is refused and nothing runs.
  • Pick its access. Read only lets it search, count and read, and change nothing. Enriching, building lists and running campaigns need those products, or full access. See Scopes.

Find an audience

Start by asking how many, then ask for the people:
Counting is free. Listing costs $0.001 for each person returned, so 20 people cost $0.02. The assistant tells you what the call charged. A description is turned into filters. Ask the assistant which filters it used: when part of your description could not be matched, the search is broader than you meant.

Get their contact details

A work email costs $0.02 when one is found and nothing when none is. A phone number costs $0.05. Asking for the wider search, the waterfall, finds more and costs $0.05 an email and $0.10 a phone number when it is the one that finds them. Enriching a list runs as a job. The assistant reads the job until it ends, then tells you how many emails were found and what the run cost.

Draft a campaign

A new campaign is a draft, so nothing is sent while you edit it. Ask for changes in plain words: a shorter subject, a different opening line, a third step.
Launching is the one request that makes a campaign send. The assistant checks that the campaign is ready, tells you who it will write to, and asks you to confirm before it launches.

Follow the replies

Reading conversations is free. An app with access to campaigns can also mark them read, star them and archive them. Answering is not offered as a tool yet. Reply in the console, or with the API: see Work the inbox.

Watch the spend

The assistant can read your balance, the prices, your usage by day and every call of the last 30 days. The console shows the same under Logs, and each connected app’s spend for the day under MCP.

Requests that work better

  • Ask for a count first. It is free, and it tells you whether the audience is 40 people or 40,000 before you pay to list any of it.
  • Give a budget. “Spend at most $2” gives the assistant a number to stop at. The daily cap is what enforces it.
  • Name things. The assistant can look up your lists, campaigns and saved searches, so “the Fintech list” is enough once it exists.
  • Ask to see it first. “Show me before you launch” and “show me the list before you enrich it” keep you in the loop.
  • Do one send at a time. A request that is sent again runs again, so do not repeat a launch you are unsure about. Ask whether it went through.

What the assistant checks with you

The server asks every app to confirm with you before a tool that spends, sends or deletes. Apps that ask for approval tool by tool know which ones these are, because every tool is marked as read-only or not. An assistant is also told to treat text inside profiles, list rows and messages as data. If a reply or a profile contains instructions, the assistant should not act on them. These are instructions to the app, not guarantees. The limits you set are what hold: the app’s access, and its daily cap.