The ad library MCP server that gives Claude only winning ads

• search_ads

  Search winning ads

  Find ads by niche, by any of the ten tags, by days live, or by a line from the headline, the body copy or the video transcript. Each result comes back with the brand, the opening line, the format, the estimated spend, days live and variants.

  problem-aware UGC skincare ads that open with a number

• get_ad

  Read one ad in full

  The full record of one ad: copy, script, tags, signals, days live and variants. This is what Claude reads before it quotes a hook or maps a script.

  split this transcript into hook, problem, proof and CTA

• winning_patterns

  See what wins more often

  Which hooks, angles, formats and proof levels reach the top tier more often than the rest of the board, with the number of ads behind each one.

  which hook patterns beat the baseline in video?

• advertiser_hooks

  Read a brand's playbook

  The hook and angle mix one advertiser is running this month, so you see the pattern behind ten ads instead of guessing from one.

  what is this brand repeating across its ads?

• get_swipe_prompt

  Turn a winner into a brief

  A ready brief that maps a winning ad's structure and adapts it to the product you describe. The draft keeps the order of the argument, not the words.

  remake this ad as a 30-second UGC script for my product

• list_taxonomy

  The valid values of the ten tag dimensions, so Claude filters with exact words instead of guessing. Calling it first is the easiest way to avoid an empty search.

  list the valid niches and hook patterns

The ten dimensions are hook pattern, hook mechanism, narrative angle, reader awareness, proof level, production format, video type, landing page type, niche and CTA type. How each one is defined is in the methodology, and the design of the taxonomy in AI ad analysis.

You need an active Ad Radar plan. The MCP server is included in Membership and Agency, with no add-on.

1. Find winning oral care ads that have run 90 days or more with at least two variants. Table: brand, opening line, estimated spend, days live.

   • list_taxonomy
   • search_ads

   What comes back

   A table of the ads brands have kept feeding. Of the 52 oral care ads on the board, 33 had passed 90 days and 11 ran two or more variants. Claude also points out that the niche mixes human and pet dental care, so one follow-up drops the pet brands.

2. Open the LIVFRESH ad that starts "This new toothpaste gel formula" and split the transcript into beats.

   • get_ad

   What comes back

   The 13-line transcript, then a beat map: hook, problem, mechanism, a five-line proof stack, urgency, and a CTA that asks for a read instead of a purchase. Each beat is quoted, so you can check it against the ad.

3. Which hook patterns, proof levels and awareness levels reach the top tier more often than the board? Flag anything with fewer than 20 ads.

   • winning_patterns

   What comes back

   Tag values ranked against the board baseline of about 10%. Strong proof reached the top tier 16.3% of the time and specific-number openers 12.9%, while unaware openers sat just below average at 8.6%. That changes the brief: keep the structure, add proof you can show on camera.

4. Remake the LIVFRESH ad for my toothpaste tablets: fluoride-free, 60-day jar, $24. UGC, 30 seconds.

   • get_swipe_prompt

   What comes back

   A structure map of the donor ad, a first draft for your product at about the same length, and a side-by-side showing where the draft keeps the structure and where it changes the content. You still replace any claim you can't back for your own product.

The full session, tool call by tool call, is in Claude for marketing research: a competitor ad walkthrough. For more questions to ask, see 25 Claude prompts for Facebook ads and the swipe file AI workflow.