Plateop vehicle check

plateop-check-mcp

A remote MCP server (Streamable HTTP, stateless) that gives AI agents the free facts about a used car from Plateop: register data, inspection due date, odometer verdict, insurance, inspection findings and recalls, each with source and as-of date. It is a thin layer over Plateop's public JSON channel (GET /api/rapport/kenteken/{plate}, GET /open/model/<brand>/<model>.json). It has no data layer, no cache, no database and no accounts.

Revenue model: every vehicle_check answer ends with a link for a person to the paid report (4.95 euro in the Netherlands) on the country host, https://<host>.plateop.com/analyse/nieuw, without the plate in the URL. An Authorization header sent by the agent is passed to the app unchanged, so a paid tier later needs no change here.

Source: https://github.com/gvb9/plateop-check-mcp. Deployment: ./server-setup.sh once, then ./deploy.sh. Published in the MCP Registry as com.plateop/plateop-check 0.1.0 (18 Oct 2026); a new version needs a version bump, mcp-publisher login http and mcp-publisher publish (see agent-commons docs/VINDBAARHEID.md).

Tools

Countries: nl (host nederland), gb (uk), il (israel), is (island), dk (danmark). Each call sends Host and X-Forwarded-Host: <host>.plateop.com (the app reads Host; node:http is used because fetch cannot set it), X-Real-IP and X-Forwarded-For with the agent's IP, and the agent's Authorization header if there is one. Timeout 10 s. App errors (400, 404, 429 with Retry-After, 502, 503) become one plain sentence for the agent.

Routes: POST /mcp, GET / (text), /llms.txt (guide), /health, /privacy, /icon.svg, /.well-known/mcp/server-card.json. server.json is for the MCP Registry (com.plateop/plateop-check).

Left out on purpose: compare_models. The app's /open/vergelijk/<pair>.json only answers for a short frozen list of model pairs on nl, gb and cz, with no way to list them, so a tool would mostly return "not found". Add it when agents ask for it.

Limits (per hashed IP, in memory)

vehicle_check 20 per 10 minutes and 100 per day (the same as the app's own limit); model tool 60 per minute; plain GET routes 600 per hour. No timers, no disk: stale windows are pruned on use and everything resets on restart.

Environment variables

There is no DATA_DIR: nothing is written to disk.

Run

npm install
npm test        # vitest against a fake Plateop (test/fakePlateop.ts); no real calls
npm run build
PLATEOP_BASE_URL=http://127.0.0.1:3105 TRUST_PROXY=1 npm start

Rollout (nothing of this has been done)

1. On the server: useradd --system plateop-check, mkdir -p /srv/plateop-check /etc/plateop-check, env file /etc/plateop-check/plateop-check.env (root:plateop-check, mode 640) with PORT=3200, PLATEOP_BASE_URL=http://127.0.0.1:3105, TRUST_PROXY=1.
2. systemd unit plateop-check (like commons): ExecStart=/usr/bin/node /srv/plateop-check/current/dist/index.js, EnvironmentFile=, User=plateop-check, Restart=on-failure.
3. nginx site zz-plateop-check for check.plateop.com behind Cloudflare, like zz-commons: proxy_pass http://127.0.0.1:3200, proxy_set_header X-Real-IP $remote_addr (or the Cloudflare IP module's value), X-Forwarded-Proto, Host. No DNS change: *.plateop.com already points at the server.
4. The privacy page promises that the journal keeps the call log for at most 30 days: set MaxRetentionSec=30day in /etc/systemd/journald.conf (or change the sentence in src/privacy.ts).
5. ./deploy.sh, then curl https://check.plateop.com/health, then one vehicle_check with an MCP client.
6. Publish server.json to the MCP Registry under com.plateop.
7. After 30 days count calls per tool and per shortened hash: journalctl -u plateop-check | grep tool=.