MCP (ChatGPT / Claude) · Cosmetics contract manufacturer

Cosmetics contract manufacturer → indie beauty brands in France: what one real Prosp run returned

By the Prosp team · run on 2026-10-03 in production

Run date

2026-10-03

Job duration

14 min 13 s

Results came from

Live web search

Delivered

8 of 8 requested

Rejected candidates

60

The scenario

A cosmetics contract manufacturer wants small, independent French beauty and skincare brands that could outsource production to it. The buyer is a profile we set up to test this market, not a customer. The run is real, and every company on this page came back from it.

Buyer

A manufacturer that formulates and fills skincare and cosmetics for other brands. An illustrative profile, not a named customer.

Target

Micro and small independent beauty and skincare brands in France.

The exact request

Sent by an AI agent through the MCP server as the tool call find_prospects with these arguments:

{
  "criteria": {
    "query": "Find French independent beauty and skincare brands that could outsource product manufacturing to a cosmetics contract manufacturer",
    "industries": [
      "cosmetics_beauty"
    ],
    "locations": [
      "France"
    ],
    "company_sizes": [
      "micro",
      "small"
    ],
    "keywords": [
      "indie brand",
      "skincare",
      "clean beauty",
      "marque"
    ],
    "maxResults": 8
  }
}

How this run was actually started: through Prosp's internal job endpoint, with exactly these criteria and a target of 8 companies, under an internal exploration account whose results never reach a CRM. That path does not charge credits. Through MCP, an agent sends these arguments to find_prospects, then calls get_discovery_status and get_discovery_results with the job id it gets back.

What came back

Checking the registry ids

Why candidates were rejected

What went wrong

• Nuxe is a large, established French brand, not an indie one. Its company_size is empty, so the micro/small preference could not exclude it.
• Avril came back as medium-sized (100 employees, a language-model estimate) under a micro/small request. Company size is a preference, not a hard filter.
• 6 of 8 rows have no registry id. For Les Huilettes the register match was not confirmed; for the other five the lookup found nothing.

How to read this run

This was the slowest of the six runs in this series. icp_intent_mismatch (15) leads the rejection histogram, ahead of content_thin (8). No candidates were discarded after the target was reached.

Six of the eight are the kind of company the request described: small French skincare brands selling under their own name (Patyka, VELAVI COSMETICS, Oolution, Les Huilettes, MATIERE BRUTE LAB, BIVOUAK). MATIERE BRUTE LAB is registered under a perfume and toiletries manufacturing code, and its employee count (3) comes from the register. Nuxe and Avril are the exceptions, covered under "What went wrong".

What this market can and can't verify

• Registers: yes, free and fast. France's SIRENE company register can be searched through the public recherche-entreprises.api.gouv.fr service. On 2026-10-03 we looked up both SIREN numbers Prosp delivered. Both matched on name and seat and were active. Each lookup took under a second; the links in the table above run the same search.
• Gaps: six rows came back without a SIREN. That is the larger problem in this market: a check that takes a second is only useful on the rows that carry an id.

Cost and time

• The job took 14 min 13 s. At submission the API estimated a median of 565 s and a 90th percentile of 1,686 s.
• A discovery costs 100 credits through the REST API, the MCP server or the Telegram bot. This run was started on an internal account and was not charged.

How to reproduce this run

• ChatGPT or Claude: connect the MCP server and ask your assistant for the same companies, or have it call find_prospects with the arguments above.
• REST API: send the same criteria to POST /api/v1/discoveries. See the API reference.
• Telegram: run /find in the bot. See the Telegram guide.

A repeat will not return exactly these eight. The web changes, and each run adds to Prosp's knowledge base, so a later run may be served partly from companies found by earlier ones.