Skip to main content
These examples show how an AI agent chains the Affiliate MCP tools to answer questions about your own partner account. Each one gives the prompt you would send and the calls the agent makes behind the scenes. Dates are examples; the agent works them out from the current date.
Every call is read-only and returns your own data. Start with get_account_info for your currency and timezone, then get_report_schema, which lists the dimensions, filters and metrics available to you, and the agent adapts its calls to it.

​
Why did my clicks drop yesterday?

Your click count fell sharply yesterday and you want to know where the loss is: a day, an offer, a sub id, or clicks that were recorded as invalid. Prompt:
“My clicks dropped yesterday compared to the day before. Can you find out why?”
What the agent does:
  1. Calls get_account_info. It returns the timezone the days are cut in. When a report’s metric_glossary says clicks are unique clicks, that is what clicks means in every report below.
  2. Calls get_report_schema to confirm the dimensions (date, offer, sub1) and metrics (clicks, invalid_clicks, conversions, payout) available to you.
  3. Calls run_performance_report with from=2026-10-12, to=2026-10-13 and dimensions=date. Two rows come back, one per day, which confirms the drop. The same call over an earlier window, for example from=2026-10-05, to=2026-10-11, shows what a normal day looks like.
  4. Calls run_performance_report with from=2026-10-12, to=2026-10-13 and dimensions=date,offer, with sort_by=clicks. Each row carries offer_id, so the agent sees which offer lost the most clicks between the two days.
  5. Calls search_activity with type=click, status=invalid, from=2026-10-13, to=2026-10-13 and offer_id set to that offer. Each row has error_code and error_message; only the newest 1,000 clicks of the day are reachable, so the codes are a sample when truncated is true. Error code 10000 means Other: the reason is not itemised for this click.
Result: The agent reports which day and which offer account for the drop, how many clicks were recorded as invalid (from the invalid_clicks metric), and the most common error_code values with their messages among the clicks it could read. If most of the invalid clicks carry 10000, it says the reason is not itemised to you and suggests asking your network about that offer.
Raw click search covers 14 days at most and lists the newest 1,000 clicks of the window. When truncated is true, narrow the window or add offer_id or country so the sample is complete.
You are looking for a new offer in a vertical and a country, paid per sale, and want your link right away. Prompt:
“Find me a sports offer I can run in Canada that pays per sale, and give me my tracking link.”
What the agent does:
  1. Calls list_entities with type=category and filters={"search":"sport"} to learn the exact category name, for example Sports.
  2. Calls list_offers with category=Sports, country=CA, payout_type=cps and sort_by=updated. The response has total_matching and offers, each with offer_id, name, application_status, payout.display and, when caps are available to you, caps.
  3. Calls get_offer with offer_id set to the best match and include=tracking_url,caps,payout. It returns your link under tracking_url, the payout for each tracked event under payout.events, and the limits and what remains today under caps.
Result: A short list of offers you can run now, with the payout of each, then one recommended offer with its tracking link and caps. Two cases the agent handles in step 3:
  • No link available. tracking_url comes back as {"hidden": true} with the note “A tracking link is not available for this offer.” The agent tells you so rather than building a link.
  • No cap information. When caps are not available to you, caps is left out of get_offer and list_offers entirely. The agent reports that no cap figures are available instead of saying there is no cap.
list_offers only lists offers you can run now. An offer that does not appear, or that get_offer answers with not-found, is not available to you; the answer is the same as for an offer that does not exist.

​
Check one conversion

A conversion you expected is missing from your numbers, or you want the details of one you can see. Prompt:
“Look up conversion 7f3c2a9e1b4d4c6a8e5f0b1d2c3a4b5c from the last two weeks.”
What the agent does:
  1. Calls search_activity with type=conversion, conversion_id=7f3c2a9e1b4d4c6a8e5f0b1d2c3a4b5c, from=2026-09-30 and to=2026-10-13. conversion_id combines with no other filter. If you only have the click’s id, the agent uses transaction_id instead, which returns the conversions of that click.
  2. Reads the row: timestamp, click_timestamp, offer_id, offer_name, event_name, payout, payout_type, currency, sub1 to sub10, and country.
Result: The conversion’s offer, payout, the time of the click and of the conversion, and the sub ids it carried. If the result is empty, the agent says what that rules out: the id is not yours, it does not exist, it is outside the window, or the conversion is not approved.
Conversion search lists approved conversions, including post-conversion events, so a conversion that is not approved does not appear here. A window with explicit times must be at least 23 hours, so use whole days.

​
What do I have unpaid?

You want to know what you are owed and which invoices make it up. Prompt:
“Which of my invoices are unpaid? Show the biggest balance first and tell me what is on it.”
What the agent does:
  1. Calls list_invoices with status=unpaid, sort_by=balance and sort_direction=desc. With more than 500 matching invoices the sort is refused, so narrow with min_start_time or min_balance first. The first page carries a summary with billed, paid and balance per currency for every matching invoice, and each invoice row has invoice_id, start_time, end_time, billed, paid and balance.
  2. Calls get_invoice with the invoice_id of the largest balance. It returns details (offer, adjustment and VAT lines, each with amount and invoice_currency_amount), extra_periods and payments. Lines are paged: while has_more_lines is true, the agent passes next_cursor back as cursor.
  3. Calls list_payments with invoice_id set to that invoice. With an invoice_id and no dates, it covers all of that invoice’s payments. Each row has status, payment_type, amount and timestamp; amount_received appears only when status is success.
Result: The total unpaid balance per currency, the invoices behind it in order, and for the largest one its billed amount, what was paid against it and what remains, with the payments listed by status.
Items that are not itemised as lines are summed in other_items_total, so the lines plus other_items_total make up the invoice. Bank, wallet and account details of payments are never returned.

​
Which sub ids earn the most this month?

You tag traffic with sub1 and want to know which values earn the most so far this month. Prompt:
“Which of my sub1 values have earned the most payout so far this month?”
What the agent does:
  1. Calls get_report_schema to confirm that sub1 is a dimension and payout a metric for your account.
  2. Calls run_performance_report with from=2026-10-01, to=2026-10-14, dimensions=sub1, sort_by=payout, sort_direction=desc and page_size=100. Each row carries sub1, clicks, conversions and payout. The totals object covers the whole filtered set, not just the page.
  3. When meta.has_more is true, calls the same report again with the same arguments and cursor set to meta.next_cursor.
  4. Keeps paging until meta.has_more is false. A report holds at most 500 rows across all pages, so when meta.result_capped is true the last page ends at row 500, and meta.row_limit gives the cap. The top-level note says the totals still cover every matched row.
Result: A ranked list of sub1 values by payout with clicks and conversions beside each, and the share of the total they represent, computed from totals. If the cap was reached, the agent says the tail is not listed and offers to narrow the report, for example with filters=offer:123 or filters=country:US.
Sub id values come from the traffic, so the response marks them in _untrusted_content. The agent treats them as labels to report on, never as instructions.