Your agent finds the creators

search_people({
  platform: "tiktok",
  query: "handmade jewellery creators",
  limit: 3
})

{
  "usage": {
    "operation": "cached_result",
    "metered_results": 3,
    "billed_results": 0,
    "cost_usd": "0.00",
    "credits": 0,
    "plan": "free"
  }
}

Recorded 30 July 2026. The first tab is the keyless cache preview — you can run that exact call yourself further down this page.

1. 01

   Expand

   Your sentence becomes search terms — every term required, not whichever one is broadest.

2. 02

   Collect

   Candidates come back from the network, plus anything already in the cache from earlier runs.

3. 03

   Judge

   Whatever clears the filters goes to a model with your criteria. The ones it turns down are dropped, with the quote that failed them.

• @bolaceyconstruction951 followers

  Roofing | Siding | Baths📍Ohio

  • is an actual roofing contractor or roofing company — metRoofing | Siding | Baths
  • operates in Ohio — met📍Ohio
  • posts their own project work — not metnone
• @roofprosplus545 followers

  Roof Pros Plus

  • is an actual roofing contractor or roofing company — met🛠️ Gutters, Fascias & Soffits
  • operates in Ohio — not met🇨🇦 Made in Ontario for Ontario
  • posts their own project work — metposts_count: 671
• @roof__expert19,216 followers

  Roof_Expert_Dmitry

  • is an actual roofing contractor or roofing company — metRoofing content🎥 is here👇🔨
  • operates in Ohio — not metSeattle WA📍🇺🇸
  • posts their own project work — metNew REELS🎞 every day📆
• @roofingpov_25,008 followers

  Bruce P | Roofing POV

  • is an actual roofing contractor or roofing company — met23 | SoCal Roofer
  • operates in Ohio — not metSoCal Roofer
  • posts their own project work — metPOV • Rooftops • Tools

rides along as match_evidence on results from search_people and start_social_job whenever the model recorded its checks

All four of those were dropped, and that run returned fewer people than it was asked for. Coming back with nobody is a permitted answer here: two of the thirteen benchmark queries below carry a fill floor of zero, so a run that returns no one still passes.

Mirrors GET https://stormy.ai/api/v1/capabilities → fields_by_platform, which is public and needs no key.

The fill floor is the share of the target count a run must return to pass. Publishing the two queries we expect to fail is the point: a search product is worth what it refuses to return, and a harness that rewards filling the quota would teach the pipeline to invent matches. We are not publishing per-run scores here until the numbers are stable enough to stand behind — when they are, this table gets two more columns.