Parts Finder MCP

Read-only MCP server for finding, comparing, and shortlisting engineering parts from distributor and marketplace APIs.

Parts Finder is designed for Codex, Claude, and other MCP-capable agents that help engineers identify embedded parts, connectors, cables, motors, sensors, PLC components, and related hardware from rough text or image-derived clues.

The project is MIT licensed. Supplier data remains subject to each provider's API terms.

Status

Implemented:

• Mouser Search API keyword search

• DigiKey ProductInformation V4 keyword search

• AliExpress Open Platform signed read-only marketplace search adapter

• in-memory TTL cache for repeated supplier searches

• candidate confidence summaries and verification checklists

• bounded multi-query expansion from category and visual hints

• Korean and field-language query normalization for common part terms

• automatic connector/motor hint inference from rough query text

• cross-supplier candidate ranking, de-duplication, and constraint filtering

• live lookup, comparison, alternate suggestion, and BOM enrichment flows

• Image-observation normalization through extract_visual_part_hints

• Local rate-limit metadata and guardrails

• Codex plugin wrapper and skill

• Smoke tests for Mouser, DigiKey, and MCP stdio execution

Planned:

• supplier-specific datasheet and lifecycle enrichment

• deeper electrical/mechanical parameter extraction from supplier attributes

• stronger alternate-part scoring from package, pinout, ratings, and lifecycle data

Related MCP server: ls-apis-mcp

Process Flow

flowchart TD
  A["Engineer asks for a part or uploads a part photo"] --> B["Codex/Claude extracts visual and text observations"]
  B --> C["MCP tool: extract_visual_part_hints"]
  C --> D["Normalize hints: markings, pin count, pitch, color, keying, board context"]
  D --> E["MCP tool: search_parts"]
  E --> F["Supplier adapters"]
  F --> G["Mouser Search API"]
  F --> H["DigiKey ProductInformation V4"]
  F --> I["AliExpress Open Platform marketplace search"]
  G --> J["Normalize candidates"]
  H --> J
  I --> J
  J --> K["Merge duplicate MPNs across suppliers"]
  K --> M["Score by exact match, visual hints, constraints, stock, price, lifecycle"]
  M --> L["Agent explains likely matches and verification steps"]

Image-Based Search Flow

This MCP server does not process raw images directly. The agent should inspect the image first, then pass structured observations into the MCP tool.

Good visual observations include:

• visible text, logos, top markings, labels, and terminal legends

• connector pin count, row count, pitch, color, shroud, key slot, latch, and wire count

• IC package shape, pin count, board position, nearby crystal/regulator/driver/interface parts

• motor diameter, length, shaft type, gearhead, encoder, cable count, and connector type

• cable jacket markings, connector gender, locking style, and estimated length

• scale clues such as USB ports, 2.54 mm headers, terminal blocks, screws, rulers, or known boards

Example:

{
  "userGoal": "identify black shrouded connector on STM32 development board photo",
  "visibleText": ["3V3", "GND", "RST", "A0", "Mini USB"],
  "packageShape": "black shrouded dual-row box header with key slot",
  "connectorPinCount": 20,
  "connectorRowCount": 2,
  "connectorPitchMm": 2.54,
  "connectorFamily": "IDC box header",
  "connectorMountingStyle": "through hole",
  "boardContext": [
    "STM32 style development board",
    "JTAG or SWD debug/programming connector",
    "2x10 IDC ribbon cable header"
  ]
}

The resulting search should be verified against dimensions, datasheets, pinout, and mating connector before purchasing.

Tools

• search_parts: search configured suppliers for engineering part candidates.

• extract_visual_part_hints: turn image-recognition observations into searchable part hints.

• lookup_part: look up a known manufacturer or supplier part number.

• compare_parts: compare known candidates.

• suggest_alternates: suggest alternate parts with caveats.

• enrich_bom: enrich BOM-like rows.

Version 0.1 is read-only. It must not place orders, create carts, call dropshipping/order APIs, or mutate supplier accounts.

Search Quality Behavior

The MCP now improves rough searches before returning results:

• builds up to four bounded query variants from the original query, category hint, visual hints, and exact-looking part numbers

• normalizes common Korean/field terms such as 2핀, 커넥터, 패널마운트, 방수, 기어모터, 엔코더, 단자대, 푸시인, and PLC 입출력 모듈 into supplier-friendly English search terms

• normalizes common passive, power, and semiconductor field terms such as 저항, 콘덴서, 인덕터, LDO, DC-DC, MOSFET, 다이오드, 10 키로 옴, 100 나노 패럿, and package sizes like 0603

• prioritizes supplier-friendly normalized queries when field-language input was translated, while still preserving the original query for traceability

• adds compact supplier queries from inferred visual hints, such as connector family, pin count, pitch, gender, and mounting style

• runs up to two relaxed fallback queries only when the initial ranked result set is empty, limiting extra API calls while improving recovery from over-specific searches

• joins split OCR-like part-number tokens such as STM32 C552 RET6 into exact supplier queries like STM32C552RET6

• uses visible OCR text from images as search queries and ranking evidence, including joined top-marking part numbers

• uses motor image hints such as shaft diameter, body diameter, body length, encoder, gearhead, and connector type in search expansion and ranking

• uses general package dimensions such as length, width, and height in visual search expansion, ranking evidence, and verification checklists

• scores candidates against those normalized terms too, so field-language searches can match English supplier descriptions more reliably

• infers pin count, row count, pitch, connector family, mounting style, color, wire count, and motor encoder/gearhead hints from rough query text

• returns per-candidate confidence, fitSummary, and verificationChecklist fields to make sourcing decisions easier to audit

• ranks exact manufacturer or supplier part-number matches above loose keyword matches

• merges duplicate candidates that share the same normalized manufacturer part number

• filters hard constraints such as manufacturer, required terms, forbidden terms, max unit price, max MOQ, RoHS, and stock

• extracts supplier attributes/parameters into normalized specs when available

• boosts useful sourcing evidence such as stock, datasheet, product URL, pricing, visual-hint matches, pin count, row count, pitch, mounting style, and connector family

• lowers confidence for obsolete/discontinued lifecycle text and marketplace results unless marketplace use is explicitly allowed

The same ranked search foundation powers lookup_part, compare_parts, suggest_alternates, and enrich_bom.

Requirements

• Node.js 20 or newer

• npm

• API credentials for any supplier you want to enable

Install

npm install
npm run build

Run Locally

npm run dev

For built stdio execution:

npm run build
npm start

Without API keys, the MCP server still starts and reports skipped-supplier status. This lets Codex, Claude, or MCP Inspector validate tool wiring before credentials exist.

Environment Setup

Copy .env.example to .env.

cp .env.example .env

On Windows PowerShell:

Copy-Item .env.example .env

Never commit .env. Keep all supplier secrets local.

Common Settings

PARTS_FINDER_LOG_LEVEL=info
PARTS_FINDER_CACHE_DIR=.cache/parts-finder
PARTS_FINDER_CACHE_TTL_SECONDS=300
PARTS_FINDER_DEFAULT_COUNTRY=US
PARTS_FINDER_DEFAULT_LANGUAGE=en
PARTS_FINDER_DEFAULT_CURRENCY=USD

PARTS_FINDER_CACHE_TTL_SECONDS controls the in-memory supplier search cache. The cache reduces repeated API calls across expanded queries, lookup_part, compare_parts, and BOM enrichment while the MCP server process is running. Set it to 0 to disable caching. PARTS_FINDER_CACHE_DIR is reserved for future persistent cache support.

Mouser

Required:

MOUSER_SEARCH_API_KEY=
MOUSER_API_BASE_URL=https://api.mouser.com

Optional local guardrails:

MOUSER_RATE_LIMIT_PER_MINUTE=30
MOUSER_RATE_LIMIT_PER_DAY=1000

Get a Mouser Search API key from the Mouser API Hub. The current adapter uses keyword search and normalizes returned candidates into the MCP part schema.

DigiKey

Required:

DIGIKEY_CLIENT_ID=
DIGIKEY_CLIENT_SECRET=
DIGIKEY_LOCALE_SITE=US
DIGIKEY_LOCALE_LANGUAGE=en
DIGIKEY_LOCALE_CURRENCY=USD
DIGIKEY_CUSTOMER_ID=0

Production:

DIGIKEY_SANDBOX=false
DIGIKEY_API_BASE_URL=https://api.digikey.com
DIGIKEY_TOKEN_URL=https://api.digikey.com/v1/oauth2/token

Sandbox:

DIGIKEY_SANDBOX=true
DIGIKEY_SANDBOX_API_BASE_URL=https://sandbox-api.digikey.com
DIGIKEY_TOKEN_URL=https://sandbox-api.digikey.com/v1/oauth2/token

Important DigiKey notes:

• ProductInformation V4 uses OAuth 2.0 client credentials for this MCP use case.

• The app must be authorized for ProductInformation V4 in the same environment as the token endpoint.

• Production credentials work with api.digikey.com; sandbox credentials work with sandbox-api.digikey.com.

• If the token endpoint returns 401 Invalid clientId, the Client ID is invalid for that endpoint/environment.

• If OAuth succeeds but product search returns 403, the app is likely not authorized or subscribed to the requested ProductInformation API.

• The DigiKey portal may ask for an OAuth callback URL. For this client-credentials flow it is not used at runtime; https://localhost is acceptable as a placeholder if the field is required.

Optional local guardrails:

DIGIKEY_PRODUCT_INFORMATION_RATE_LIMIT_PER_MINUTE=120
DIGIKEY_PRODUCT_INFORMATION_RATE_LIMIT_PER_DAY=1000

AliExpress

AliExpress support is read-only and only participates in search_parts when constraints.marketplaceAllowed is true. This prevents marketplace listings from silently mixing into production distributor recommendations.

The Open Platform exposes different product-search APIs depending on app approval and product line. Set ALIEXPRESS_PRODUCT_SEARCH_PATH to the search API path granted to your app. The default is a dropshipping-style text search path.

ALIEXPRESS_APP_KEY=
ALIEXPRESS_APP_SECRET=
ALIEXPRESS_ACCESS_TOKEN=
ALIEXPRESS_REFRESH_TOKEN=
ALIEXPRESS_API_BASE_URL=https://api-sg.aliexpress.com
ALIEXPRESS_PRODUCT_SEARCH_PATH=/aliexpress/ds/textsearch
ALIEXPRESS_OAUTH_AUTHORIZE_URL=https://api-sg.aliexpress.com/oauth/authorize
ALIEXPRESS_TOKEN_CREATE_PATH=/auth/token/security/create
ALIEXPRESS_TOKEN_REFRESH_PATH=/auth/token/refresh
ALIEXPRESS_COUNTRY=US
ALIEXPRESS_CURRENCY=USD
ALIEXPRESS_LANGUAGE=en_US
ALIEXPRESS_RATE_LIMIT_PER_SECOND=
ALIEXPRESS_RATE_LIMIT_PER_DAY=

AliExpress rate limits vary by app key, API, and app-key/API combination. Copy the approved quotas from the Open Platform console after the app is reviewed.

Marketplace caveat: AliExpress results should be treated as prototyping or long-tail sourcing leads unless the exact seller, variant, dimensions, authenticity, shipping terms, and ratings are verified.

MCP Configuration

After npm run build, add this server to an MCP-capable client.

{
  "mcpServers": {
    "parts-finder": {
      "command": "node",
      "args": ["C:/absolute/path/to/parts-finder-mcp/dist/index.js"]
    }
  }
}

If the client does not load .env from the project directory, pass environment variables explicitly in the MCP config or launch the client from the project root.

Codex Plugin

The Codex plugin wrapper lives at:

plugins/codex/parts-finder

It contains:

• .codex-plugin/plugin.json

• .mcp.json

• skills/parts-sourcing/SKILL.md

Validate it with:

python C:/Users/googo/.codex/skills/.system/plugin-creator/scripts/validate_plugin.py plugins/codex/parts-finder

The plugin .mcp.json points to the built MCP server:

{
  "mcpServers": {
    "parts-finder": {
      "command": "node",
      "args": ["../../../dist/index.js"],
      "env": {
        "PARTS_FINDER_LOG_LEVEL": "info"
      }
    }
  }
}

Smoke Tests

Build first:

npm run build

Run supplier smoke tests:

npm run smoke:mouser
npm run smoke:digikey

Run MCP stdio smoke tests:

npm run smoke:mcp
npm run smoke:mcp-cache
npm run smoke:mcp-fit-summary
npm run smoke:mcp-digikey
npm run smoke:mcp-aliexpress
npm run smoke:mcp-workflows

DigiKey production override example:

$env:DIGIKEY_SANDBOX='false'
$env:DIGIKEY_TOKEN_URL='https://api.digikey.com/v1/oauth2/token'
npm run smoke:mcp-digikey

Expected successful DigiKey MCP smoke output for a connector query looks like:

Raw count: 1196
Candidates: 2

Rate Limits

The server stores supplier rate-limit metadata so adapters can throttle before calling APIs:

• Mouser default: 30/min, 1000/day; verify against current account/API behavior.

• DigiKey ProductInformation standard quota: 120/min, 1000/day; honor X-RateLimit-*, X-BurstLimit-*, and Retry-After headers.

• AliExpress: app-key/API/API+app-key quotas vary and should be copied from the Open Platform console after approval.

Candidate Verification Checklist

Before recommending a final part, the agent should verify:

• exact manufacturer part number

• supplier part number

• pin count, pitch, rows, gender, keying, latch, and mounting style

• voltage, current, temperature, and environmental ratings

• stock quantity and lead time

• lifecycle status and replacement notice

• datasheet or official product-page match

• mating connector and crimp terminal compatibility

Image-derived matches should be treated as low confidence until measurements or exact markings confirm them.

Development

npm run build
npm test
npm audit

License

MIT.