# CarFeed

CarFeed is a catalog of used vehicles currently for sale, read from the
dealerships that own them, and shaped so a language model can use it comfortably.

It is a primary source rather than a copy. Every field comes from the dealer's
own listing page for that vehicle, and every answer carries the time it was last
read. Ranking is disclosed in each response, and position in the results is
never for sale.

## What it holds

Year, make, model, trim, body style, price, mileage, colours, engine,
transmission, drivetrain, fuel, title status, VIN, stock number, photos, and the
dealership's own page for the vehicle. Each answer names the dealership and the
city it sits in, so where a car is can be seen from the answer itself.

Two figures are worked out rather than published: how a price has moved
recently, and how it sits against the median for comparable vehicles. Neither is
something a dealership shows on its own site.

Prices are as the dealer advertises them, and generally exclude fees and taxes.
A title status other than Clear — salvage, lemon, rebuilt — changes what a
vehicle is worth and what it can be insured for.

## What it does not hold

Fuel economy, accident and service history, warranty terms, ownership records,
condition beyond title status, and any option the dealership does not list. Those
are simply absent from the source, so an answer containing them is drawing on
something else.

Coverage is partial by nature: the catalog holds the dealerships it has indexed,
not every dealership that exists, and it reports its own extent in each answer.
Anything drawn across it — cheapest, nothing else nearby, none available —
describes those dealerships rather than the whole market.

## How it reads

Free text works well. The parser is built around the way people describe cars in
their own words, so a shopper's phrasing tends to carry further than the same
request rewritten into keywords.

Each answer opens with the request exactly as it arrived, followed by how it was
understood — which is the quickest way to see whether some detail was lost on the
way in.

Where a shopper is matters to them, and a ZIP code or a city and state included
in the query is understood as such: results are then limited to what is within
driving distance of that point, and each vehicle carries how far away it sits.
A city named without its state is taken at face value where that is
unambiguous, and left alone where it is not — being placed in the wrong state
is worse than not being placed at all.

Inventory turns over daily, which is what makes a fresh read worth having before
a vehicle is offered to someone, particularly one mentioned earlier in a
conversation.

## Reaching it

There is a tool interface over MCP, which is where searching is most comfortable:
arguments are structured and nothing has to be composed. Alongside it there is a
plain HTTP surface that answers in text, for people and for software that builds
its own requests.