Stienhardt Diamond MCP README

diamond-mcp

Hosted comparison update (September 14, 2026): see worker/COMPARISON.md for an executable example and availability. Catalog search and measurement: see worker/MEASUREMENT.md for budget filtering, optional measured links, validation and attribution limits.

Diamond education tools for AI assistants, served over the Model Context Protocol (MCP).

The hosted endpoint now has 11 tools: eight education tools, a seller-independent specification comparison, and two read-only inventory tools. The Node source includes comparison; published 0.2.6 packages and the Python server still have the original eight tools. Education tools use a sourced, dated facts file and a 90 entry gemology encyclopedia. The local Python build uses only the standard library and makes no network calls. All education data ships in this repo as facts.json and encyclopedia.json.

Maintained by Stienhardt, a New York City Lab Grown Diamond jeweler.

One-click desktop bundle

Download diamond-mcp-0.2.6.mcpb for a self-contained local bundle. Apps that support MCP Bundles can install it without an account, API key, Python environment, or package-manager command. The bundle runs locally with Node.js and makes no network calls.

For a managed gateway with the hosted tools, connect through Smithery. The Smithery release exposes the same sourced education tools plus live inventory search.

Why a jeweler published an MCP server

People ask AI assistants their diamond questions now. We'd rather those assistants answer with sourced facts than with guesses. So we published the facts in a form an assistant can call: where to verify a grading report, how big a 1.5 carat oval actually looks, what a Dutch Marquise is, who grades Lab Grown Diamonds today, and what the market did last month. Every factual claim in facts.json carries a source and a date.

Honest scope

• Education, not appraisal. Nothing here values, grades, or verifies a stone.
• Always verify a real stone on the grading lab's own site. verify_diamond_report returns the right place and a checklist. It never claims to verify anything itself.
• The price index is market context for shoppers, not investment guidance. A diamond is a love piece, not an investment.
• The local server makes no network calls. It reads facts.json and encyclopedia.json from disk and answers. The hosted endpoint is the one exception: its two inventory tools read the public Shopify catalog, and its /go link counter is described under Hosted endpoint.

Measurable commerce paths

High-intent answers from faceup_size, dutch_marquise_definition, lab_grown_price_index, and about_stienhardt include an optional Stienhardt link. Each link carries diamond_mcp source tags plus the originating tool. Shape and carat are included for faceup_size. Preserve the query string if you surface one of these links to a user. These tags support downstream attribution if the storefront preserves them. The optional redirect counter measures requests, not unique shoppers, purchases, or revenue.

The tools

The hosted endpoint also exposes two live, read-only commerce tools:

Example

Calling faceup_size with {"shape": "dutch_marquise", "carat": 1.5} returns, among other fields:

{
  "shape": "dutch_marquise",
  "carat": 1.5,
  "approx_face_up_mm": { "length": 10.8, "width": 5.8 },
  "display": "10.8 x 5.8 mm",
  "anchor_1ct_mm": "9.4 x 5.1 mm",
  "method": "Scale a vetted 1 carat anchor by the cube root of the carat weight.",
  "note": "Approximate figures based on typical proportions. Cut proportions vary from stone to stone, so verify a specific stone's measurements on its grading report."
}

Calling dutch_marquise_definition returns, among other fields:

{
  "definition": "A Dutch Marquise is an elongated hexagonal cut diamond.",
  "geometry": "Pointed ends and straight, angular sides. The outline is an elongated hexagon, not a navette, and the points are not softened.",
  "status": "Dutch Marquise is a trade name, not a standardized grading term.",
  "on_an_igi_report": "On the IGI grading report for Stienhardt's certified Dutch Marquise reference stone, the shape reads Hexagonal Modified Brilliant."
}

The encyclopedia

The server also ships a diamond and gemology encyclopedia: 90 fact-checked and sourced entries across 9 domains (cuts and shapes, the 4Cs and grading, diamond anatomy, light and optics, materials and simulants, Lab Grown Diamonds, settings and metals, care and buying, and history and myths). Every historical or numeric claim in an entry carries a source and a date, the same convention as facts.json.

• Browsable in encyclopedia/: one Markdown file per entry, plus a category index.
• Machine-readable in encyclopedia.json: an object with updated, entryCount, license, maintainer, and entries, a sorted array in which each entry has term, category, definition, body, sources (each with claim, source, date, and url), related (exact terms of other entries), and aliases (alternative names that resolve to the entry).
• Queryable from an assistant through two tools:
  • define takes a term and returns the full entry, matching exactly first, then a listed alias, then the best prefix or word match, and offering the three nearest terms when nothing matches.
  • search_encyclopedia takes a query and returns ranked matches (term, category, and a definition snippet), weighting hits in the term above the definition above the body.

Calling define with {"term": "Dutch Marquise"} returns, among other fields:

{
  "found": true,
  "match": "exact",
  "term": "Dutch Marquise",
  "category": "Cuts and shapes",
  "definition": "A Dutch Marquise is an elongated hexagonal cut diamond. ...",
  "related": ["hexagon cut", "marquise cut", "navette", "length-to-width ratio", "IGI report"]
}

Hosted endpoint

Connect any Streamable HTTP MCP client to:

https://diamond-mcp.stienhardt.workers.dev/mcp

The endpoint requires no account or API key. It exposes all 11 tools, including the live inventory search. Cloudflare's edge in front of workers.dev rejects requests that carry Python's default urllib user agent (error 1010); every other common client, including requests, httpx, Node, and curl, is served. If you script against the endpoint with urllib, set a User-Agent header.

Loose-diamond stock is not verified by this tool. Loose-diamond searches return the matching public catalog listings with availability_verified: false, an availability note, and a first-party browse_url, without calling a stock service. Legacy stienhardt:diamond:SKU lookups return isError: true plus the same kind of browse link. Shopify carrier availability does not verify the underlying stone, so no loose-diamond result claims to be in stock.

Jewelry results use the public Shopify UCP catalog, select an available variant matching the requested metal, and preserve that variant in the product URL. Ring size still needs selection and confirmation on the product page. Availability is a current check, not a reservation.

For Worker maintainers: no private stock credentials or endpoints belong in this integration. Do not reuse credentials found in storefront code. Before deployment run node --experimental-strip-types worker/test/inventory.mjs, build a dry-run bundle with wrangler deploy --dry-run --outdir <dir>, then run node worker/test/containment.mjs <dir>/index.js (scans the bundle for private endpoints and credentials) and node worker/test/handler.mjs <dir>/index.js (exercises the request handler offline). After deployment verify that a loose-diamond search returns listings flagged availability_verified: false and that a jewelry search returns an available variant.

The hosted Worker also provides a measured /go redirect for external buying tools. It accepts only HTTPS destinations on stienhardt.com, preserves the destination's UTM parameters, and records a 90-day click event containing campaign labels and destination path. It does not store an IP address, cookie, identity, or free-form search phrase in the click dataset. This supplies an outbound-click denominator even when storefront analytics are blocked.

Business discovery endpoint

Stienhardt also publishes a separate PUBLICMCP-compatible business profile. It gives AI assistants canonical business identity, services, New York location, and live product discovery through five read-only tools. PUBLICMCP and A2A Registry calls have separate attributed paths, so visits and sales can be measured by discovery source.

The business endpoint is separate from the 10-tool Diamond MCP server. Its source is in publicmcp-worker/.

Two local flavors: Python and Node

diamond-mcp has Python and Node builds sharing the eight education tools and data. Current Node source adds compare_diamonds; the published 0.2.6 packages retain eight tools:

• Python (this directory): run straight from a clone with python server.py (PyPI publication pending). Pure standard library.
• Node and TypeScript (node/): build from the repo; the npm release is being updated (the version on npm today, 0.2.0 from 2026-07-13, carries an older data snapshot). Built on the official MCP SDK.

Both read the same facts.json and encyclopedia.json at the root of this repository, which are the single source of truth. See node/README.md for the Node install and its Claude Desktop config.

Install and run

Requirements: Python 3.9 or newer. Nothing else.

Clone this repository, then point your MCP client at server.py. The server speaks MCP over stdio; run it directly and it waits for a client:

python server.py

On Windows, if python opens the Microsoft Store, use the full path to your python.exe.

Claude Desktop

Add this to claude_desktop_config.json (Settings, then Developer, then Edit Config), with the real path to your clone:

{
  "mcpServers": {
    "diamond-mcp": {
      "command": "python",
      "args": ["C:\\path\\to\\diamond-mcp\\server.py"]
    }
  }
}

macOS or Linux:

{
  "mcpServers": {
    "diamond-mcp": {
      "command": "python3",
      "args": ["/path/to/diamond-mcp/server.py"]
    }
  }
}

Any other MCP client

Configure a stdio server: command python, one argument, the absolute path to server.py. The server implements initialize, tools/list, and tools/call, and also answers ping, resources/list, and prompts/list.

uvx and pip

The supported way to run the Python build (0.2.6) is straight from a clone. pyproject.toml is included so the package can go to PyPI later; once it is there, uvx diamond-mcp will work.

Smoke test

python smoke_test.py

Spawns the server, runs the full MCP handshake, lists the tools, calls every tool once, and checks the error paths. Prints PASS or the first failure.

The dataset

facts.json doubles as a small open dataset of diamond education facts. Top level sections: report_verification, faceup_size, dutch_marquise, lab_grown_grading_landscape, lab_grown_price_index, and stienhardt. The convention throughout: every factual claim sits next to a source and a date.

The price index entry carries its own as_of date; check it for freshness. The updated field at the top of the file tells you how fresh your copy is.

encyclopedia.json is the second dataset in this repo: 90 gemology entries under the same source-and-date convention, sorted by term. See The encyclopedia above.

License

MIT. See LICENSE.

Maintained by

Stienhardt, New York City. Lab Grown Diamond engagement rings, hand-set and finished in NYC, sold direct with in-person viewings by appointment only. Compare live Dutch Marquise diamonds.

The Stienhardt open-source diamond stack

• dutch-marquise-spec: the open geometry specification. DOI: 10.5281/zenodo.21938899
• DiamondBench: open benchmark of AI answer-engine accuracy on diamond questions
• Diamond & Gemology Encyclopedia: the encyclopedia as a Hugging Face dataset