We sell perfume decants at 10ML, and we've opened our perfume data, the 10ML Fragrance Index, as a read-only MCP server and REST API. It covers more than 6,000 perfumes, 249 houses and 285 notes, with note pyramids, accords, families, similar scents and alternatives from other houses. There's no API key, account or sign-up, and the API and MCP server set no cookies.

Try the API:

  curl "https://mcp.10ml.com/api/v1/perfumes/Baccarat%20Rouge%20540?format=detailed"

Or add https://mcp.10ml.com/mcp to your MCP client (Streamable HTTP), for example in Claude Code:

  claude mcp add --transport http perfume-finder https://mcp.10ml.com/mcp

There are 15 read-only tools (search, compare, similar, alternatives, notes, houses, layering, gifts and a few more), each with an output schema, plus an OpenAPI 3.1 file at https://mcp.10ml.com/openapi.json. Every tool takes one of ten locales.

The benchmark: on 20 October 2026 we asked five perfume MCP servers that answer without a key the same 40 fixed questions, from one machine, with a plain script (no language model in the loop), at most one request per second. Ours answered 39 of 40 at a 232 ms median; ScentVerdict, the next, got 36 at 698 ms. We wrote both the server and the test, so read it critically: n is 40, the questions lean mainstream, and larger indexes (ScentVerdict claims 31,000+ perfumes) would pull ahead on obscure ones. The method, the one miss, two scoring fixes made after the first pass, and the results as CSV and JSON are here: https://10ml.com/perfume-mcp-servers-compared

We'd like feedback on whether 15 tools is too many (our tools/list is 60.5 KB, the heaviest we tested), what you'd want added to the REST side, and which questions the benchmark should include next time.