DEV Community

Scento
Scento

Posted on

Give Claude a perfumer's memory in 30 seconds

#ai #mcp #api #tutorial

Ask a language model which perfumes share a tobacco-and-vanilla base and you get a confident paragraph with a few names in it, some of which may not exist. This tutorial wires a model to a reference it can query instead, so the answer comes back with a source attached.

The reference is the Scento Perfume Compendium: 700+ perfumes, 400+ raw materials and 120+ houses, free, no key, no account. It speaks MCP, and it also speaks plain HTTP.

MCP in two lines

The Model Context Protocol lets an assistant call tools that live outside the chat. A server lists its tools, the assistant picks one, and the result lands back in the conversation as text plus structured data.

The Compendium server is remote, so there is nothing to install on your machine: one URL, https://www.scento.com/mcp.

Connect it

Claude (web or desktop). Settings → Connectors → Add custom connector → paste the URL.

ChatGPT. Settings → Apps → Advanced → turn on developer mode → Create → paste the same URL, authentication None.

Claude Code. One line:

claude mcp add --transport http scento-perfume-compendium https://www.scento.com/mcp
Enter fullscreen mode Exit fullscreen mode

Cursor, Windsurf and other mcpServers clients. Add this entry to the client's MCP config (for Cursor, ~/.cursor/mcp.json or the project's .cursor/mcp.json):

{
  "mcpServers": {
    "scento-perfume-compendium": {
      "type": "http",
      "url": "https://www.scento.com/mcp"
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

VS Code. Same entry, but .vscode/mcp.json names the top-level key servers:

{
  "servers": {
    "scento-perfume-compendium": {
      "type": "http",
      "url": "https://www.scento.com/mcp"
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Clients that only launch local commands. The same server is published as a thin proxy on npm and PyPI:

npx -y scento-perfume-compendium
# or
uvx scento-perfume-compendium
Enter fullscreen mode Exit fullscreen mode

Either one forwards to the remote URL, so you get identical answers.

What you get: twelve read-only tools

compendium-search, perfume-portrait, scent-neighbours, perfume-contrast, note-portrait, material-match, house-portrait, family-map, occasion-edit, wardrobe-composer, perfumery-lexicon and journal-search. Every tool is annotated read-only and idempotent, so a client can treat repeat calls as safe.

Every tool returns the same shape, which the server calls a folio: a quotable lede, a detail object, the scento_page it came from, further_reading links, and a cite string. That last part matters. The model has something to quote and a page to point at.

Session 1: one perfume

Prompt: "What is Tom Ford Oud Wood like, and when would I wear it?"

The assistant calls perfume-portrait:

{ "name": "perfume-portrait", "arguments": { "perfume": "Tom Ford Oud Wood" } }
Enter fullscreen mode Exit fullscreen mode

The text half of the response (live, 21 Oct 2026; the name, house and further-reading lines are trimmed):

Oud Wood by Tom Ford belongs to Scento’s Spice Route family, built on cardamom, pink pepper, patchouli and amber.

**Family:** Spice Route
**Notes:** Cardamom, Pink Pepper, Patchouli, Amber, Oud, Tonka Bean
**Character:** Smooth woods and spice: cardamom and pink pepper open into patchouli, amber and oud, warmed by tonka bean.
**Wears:** seasons: autumn, winter; hours: evening, night

Read on scento.com: https://www.scento.com/perfumes/oud-wood-tom-ford
Source: Scento Perfume Compendium, scento.com
Enter fullscreen mode Exit fullscreen mode

The model answers with the family, the notes and the autumn-evening guidance, and links the page. It did not have to guess a season.

Session 2: two perfumes side by side

Prompt: "Byredo Mojave Ghost versus Gypsy Water, what actually differs?"

{ "name": "perfume-contrast",
  "arguments": { "perfumes": ["Byredo Mojave Ghost", "Byredo Gypsy Water"] } }
Enter fullscreen mode Exit fullscreen mode

The lede, the shared materials and the response's detail.axes laid out in two columns (Mojave Ghost left, Gypsy Water right; the house, season, hour, year and concentration axes are left out):

Mojave Ghost and Gypsy Water share sandalwood, but part ways on family.

shared:    Sandalwood
family:    Wood & Root  |  Green Leaf / Amber & Resin
top:       Ambrette, Naseberry  |  Bergamot, Juniper, Lemon, Pepper
heart:     Magnolia, Sandalwood, Violet  |  Incense, Iris, Pine Needle
base:      Cedarwood, Musk, Vetiver  |  Amber, Sandalwood, Vanilla
character: woody, musky, floral, powdery  |  warm, fresh, citrusy, green
Enter fullscreen mode Exit fullscreen mode

The tool returns the axes already aligned, so the model does not have to line up two pyramids itself. A verdict array adds a one-line "suits" for each perfume, for example "Someone after something woody and musky, worn in spring, summer and autumn by day or evening." for Mojave Ghost.

Session 3: search by material, not by name

Prompt: "Which perfumes have tobacco and vanilla in the base?"

{ "name": "material-match",
  "arguments": { "materials": ["tobacco", "vanilla"], "tier": "base" } }
Enter fullscreen mode Exit fullscreen mode

The text half, with links trimmed:

5 perfumes in the Scento Compendium contain tobacco and vanilla in the base, Naxos first.

**Perfumes:**
- Naxos (Xerjoff) — family: Velvet Sweet; Tobacco base, Vanilla base
- Red Tobacco (Mancera) — family: Wood & Root; Tobacco base, Vanilla base
- Spicebomb Extreme (Viktor&Rolf) — family: Spice Route; Tobacco base, Vanilla base
- Memoir Man (Amouage) — family: Green Leaf; Tobacco base, Vanilla base
- Phi Une Rose de Kandahar (Tauer Perfumes) — family: Petal Garden; Tobacco base, Vanilla base
Enter fullscreen mode Exit fullscreen mode

materials takes one to four names that must all be present, tier accepts top, heart, base or any (the default), and family narrows the list further. If a list is longer than one page, the answer carries a continuation token to pass back in.

No MCP client? Use the REST API

The same twelve answers are available over HTTP under /api/compendium/v1, with an OpenAPI 3.1 description.

# the contrast from session 2
curl -s "https://www.scento.com/api/compendium/v1/contrast?perfumes=byredo-mojave-ghost,gypsy-water"

# the match from session 3
curl -s "https://www.scento.com/api/compendium/v1/materials/match?materials=tobacco,vanilla&tier=base"

# the spec, for codegen or a custom GPT action
curl -s https://www.scento.com/api/compendium/v1/openapi.json
Enter fullscreen mode Exit fullscreen mode

Other routes follow the same pattern: /search?q=, /perfumes/{perfume}, /perfumes/{perfume}/neighbours, /materials/{material}, /houses/{house}, /families, /edits/{occasion}, /wardrobe?loves=a,b, /lexicon/{term} and /journal?q=. Every route takes lang, and list routes take continuation. Each response is the bare folio JSON, the same object an MCP tool returns in structuredContent.

A small Python wrapper:

import requests

BASE = "https://www.scento.com/api/compendium/v1"

def match(materials, tier="base"):
    r = requests.get(f"{BASE}/materials/match",
                     params={"materials": ",".join(materials), "tier": tier},
                     timeout=20)
    r.raise_for_status()
    folio = r.json()
    return folio["lede"], folio["detail"]["perfumes"], folio["cite"]

lede, perfumes, cite = match(["tobacco", "vanilla"])
print(lede)
for p in perfumes:
    print("-", p["name"], p["house"], p["page"])
print(cite)
Enter fullscreen mode Exit fullscreen mode

Keep the cite string and the scento_page URL in whatever you render. That is the point of the exercise.

Why a tool call beats a pasted table

You could paste a spreadsheet of perfumes into a system prompt. It would go stale, eat context, and the model would still have to find the right row. A tool call does the lookup on the server and returns only the folio that matters, usually a few hundred tokens. The model decides when to call it; you do not have to route the question yourself.

It also changes how you can phrase prompts. Instead of "recall everything you know about", try "use the Compendium to compare these two and quote the cite line". Models follow that well, and the output gets easy to check: every claim has a page behind it.

If you are building your own agent loop, treat the folio as the contract. Pass lede and detail to the model, keep scento_page and cite for the UI, and use continuation when a list runs past one page.

Where the answers come from

Entries are drawn from the houses' own official pages and from authorized retailers. Each entry is traceable to a saved source, and every answer points to the Scento page that covers it. When the Compendium has no entry for something you ask about, it says so (look for the missing field in a contrast) rather than inventing one.

That is also why the answers are plain: notes, family, character and when a perfume wears best. It is a reference, not an opinion engine.

Practical notes

  • The lang argument switches the scento.com links to one of ten languages (en, de, fr, es, it, ro, el, hu, cs, pl). Answers themselves stay in English.
  • Perfume names can include the house ("Tom Ford Oud Wood") or be a slug (oud-wood-tom-ford). Include the house when the name is common.
  • Nothing needs a key, and nothing writes anything. Safe to leave enabled in an agent.

Links

Tried it with a client we did not list? Tell us which one and what the config looked like.

Scento

Top comments (0)