Ask your assistant what to check on that car
The Elgarde MCP server gives Claude — or any MCP client — a real pre-purchase inspection checklist for a specific used car, that model's documented weak points, a scoring pass, and VIN decoding. One command to install, nothing to sign up for.
What it looks like
Without it, an assistant answers a used-car question from whatever it half-remembers about the model. With it, the answer comes from a curated checklist and a defect corpus that you can read yourself.
I'm seeing a 2015 Golf 1.6 TDI tomorrow. What should I check?
33 items for this car. Beyond the usual — panel gaps, even tyre wear, complete service history — three are specific to this model and worth budgeting for: the DQ200 dry-clutch DSG, the 1.4 TSI chain tensioner, and EGR/DPF clogging on a car used mostly for short trips.
Eight of the items are marked critical — among them rust or corrosion, and whether the odometer matches the paperwork. A failure on any one of those caps the whole verdict, regardless of how well the rest scores.
Rust is fine, 4. Mileage doesn't match the service book though.
Overall 1 out of 5. The unclamped average was 2.5, but a mismatched odometer is a critical item, and a failing critical item caps the verdict. Walk away, or price it as a car with unknown mileage.
Four tools
Three of them run entirely on the machine the server is installed on — the inspection data ships inside the package. Only VIN decoding uses the network.
get_inspection_checklist
The checklist for a specific car: base items for every vehicle, fuel-specific items — DPF and glow plugs for a diesel, state-of-health and charging for an EV — and the known weak points of the model when one is curated.
EN · PT · RU · UK · FR · offline
get_known_defects
Just the curated weak points for a make and model. An empty result means nothing is curated for that model yet — the server says so explicitly, so an assistant cannot turn silence into a clean bill of health.
EN · PT · offline
score_inspection
Turns 1–5 answers into per-section scores and one verdict. A critical item scoring 2 or below clamps the verdict down to it, so a car can average 4.9 and still come out a 1.
offline
decode_vin
A 17-character VIN into make, model, year, fuel type and body style, using the free NHTSA vPIC dataset through the public Elgarde API. US-centric, so partial results on European cars are normal.
network · no key, no account
Install
Pick whichever matches your client. Nothing to configure afterwards.
Claude Code
One command, and it is available in every session.
claude mcp add elgarde -- npx -y elgarde-mcp
Claude Desktop, Cursor, and other clients that spawn a process
Add the server to the client's MCP configuration file.
{
"mcpServers": {
"elgarde": {
"command": "npx",
"args": ["-y", "elgarde-mcp"]
}
}
}
Remote — no install at all
The same four tools over streamable HTTP, for clients that connect to a URL. Stateless: the tools are read-only and keep nothing between calls.
claude mcp add --transport http elgarde https://mcp.elgarde.com/mcp
Where the answers come from
Not from a model's recollection of car forums. Every answer traces to an openly published dataset you can read, check and correct.
The inspection canon
25 base checklist items plus 10 across four fuel modules, authored in five languages, and 8 match rules covering 25 curated known defects. Published as JSON under CC-BY-4.0.
A written scoring spec
Group rollups and the critical clamp are specified, not implied — and pinned by conformance vectors that the app, the API and every language port must pass, so the same answers produce the same verdict everywhere.