Japan's largest used-car marketplace, as one JSON API

Plain yen, and the site's own string beside it so you can check us. Japanese car sites price in 万円, which is ten-thousand yen, and CarSensor prints the number and the unit in two separate places — so 68.1万円 is 681,000 JPY, not 68. The API reads the unit off the page rather than assuming it, returns price_total_jpy 681000 and price_body_jpy 561000 as integers, and returns price_total_display "68.1万円" and price_body_display "56.1万円" next to them. Mileage works identically: the card prints 1.7万km and you get mileage_km 17000 with mileage_display "1.7万km". An engine that read those numbers at face value would be out by a factor of 10,000, so we publish both and let you verify.

Because the detail page publishes the vehicle price THREE times and we return all three. It is printed as 56.1万円, carried again as an exact yen figure in the page's markup, and carried a third time in the page's own structured data. The API returns price_body_jpy (our conversion), price_body_jpy_attr and price_body_jpy_ldjson (the page's own two figures) and price_body_witnesses_agree, which compares them on every record. Measured across 12 detail records: 12 of 12 agreed, zero mismatches, on both of our two separate verification runs. We also recompute the arithmetic: price_total_matches_body_plus_fees checks that the vehicle price plus the fees equals the printed total, and that came back true 12 of 12. If any of those ever disagreed you would see it as a false in the response rather than finding out from a customer.

Because Japan quotes both, they are genuinely different numbers, and merging them would be wrong. 本体価格 is the vehicle on its own; 支払総額 is what you actually pay, including tax, registration and the dealer's 諸費用 fees. On one measured car that is 561,000 and 681,000 with 120,000 of fees in between — a 21% gap. You get price_body_jpy, price_total_jpy and price_fees_jpy as separate fields, never a single "price". Many dealers also offer optional purchase plans with their own totals, and those come back in purchase_plans: on the same car the base plan was 681,000, a roadside-assistance plan 687,000 and a cleaning plan 692,000, all on the identical vehicle price. One line against us: the vehicle price has an exact yen figure, but CarSensor publishes the TOTAL only to 0.1万円 — 1,000-yen — resolution, so price_total_is_rounded is true on every record rather than us pretending both are exact.

Every filter we expose was measured narrowing a same-run control, and we used two controls because one gives the wrong answer. The clearest case is exclude_kei: against a Prius keyword control it moved nothing at all, 11,002 to 11,002 — because no Prius is a kei car, which is an honest result and not a broken filter — and against a whole-prefecture control it bit properly, 29,908 to 20,796. Both numbers are in our published logs. 36 filters were measured this way, for example price_max 1,000,000 taking 11,002 to 2,832 and 29,908 to 8,125; year_from 2020 to 3,918 and 15,107; mileage_max 50,000 to 4,450 and 17,716; SUV to 630 and 6,175; 4WD to 952 and 4,837; seven seats to 199 and 3,412; left-hand drive to 12 and 786; two years of inspection left to 179 and 2,599. Handles the source accepts and does nothing with are deliberately not exposed, and an unknown value is rejected with INVALID_PARAM before the call is made rather than quietly returning an unfiltered result.

70 of CarSensor's own condition and equipment codes, and yes, they AND together — measured on three at once: no-accident-history alone matched 10,254 cars in a control, adding one-owner took it to 2,006 and adding a sunroof to 168, each step strictly narrower. There is no "maximum three filters" ceiling here. The list is the site's own vocabulary, covering condition (no accident-repair history, one owner, non-smoker, service-record book, registered-but-unused, ex-display, ex-rental, cold-climate spec, eco-car tax eligible), safety (autonomous emergency braking, adaptive cruise control, lane-keep assist, blind-spot monitor, 360-degree camera, parking sensors, the Japanese "Support Car" package, pedal-misapplication guard), comfort (heated and ventilated seats, leather, electric seats, three rows, walk-through, power tailgate, sunroof, dashcam, ETC toll transponder, 1500 W household outlet) and styling (alloy wheels, full aero kit, lowered or lifted suspension, resprayed). A code the source does not recognise is rejected here, because upstream it is accepted and silently ignored — which would hand you an unfiltered result that looked filtered.

Three levels. region takes one of the nine the site uses and they are a true partition — their totals summed to exactly the 519,200 national figure, so nothing is double counted or missing. prefecture takes any of the 47 by its romanised name. city takes CarSensor's own municipality code for a single town, and it composes with a prefecture: on one measured query a keyword matching 11,002 cars nationwide came down to 6 cars in one city, and that same city code under the wrong prefecture honestly returned 0 rather than silently ignoring one of them. City codes come back in the breadcrumb of every listing response, so you never have to guess one. One important line against us: a prefecture name the site does not know returns the FULL national list with HTTP 200 upstream rather than an error, so the API validates all 47 names and the 9 regions before calling and rejects a typo with INVALID_PARAM. You cannot pass a region and a prefecture together — the site has one geography slot and silently keeps one — so that combination is refused instead of guessed.

Thirty, fixed by the source, with no override — measured on pages 1, 2, 100 and 1000, with the remainder on the last page. The whole result set is pageable; there is no hidden reachable-results ceiling. But there is one trap and we handle it for you: if you ask for a page past the end, the source answers HTTP 200 with the LAST page's rows rather than an error. Measured twice — pages 17308 and 20000 of a 17,307-page set both returned page 17307's 20 rows byte for byte, and pages 368 and 9999 of a 367-page set both returned page 367's 22 rows. So every response carries last_page computed from the source's own total, page_clamped true when the source did that, and has_more which is false on the last page — follow has_more and you can never walk into it.

The default order is reproducible: two consecutive calls with no sort returned the same first five listings in the same order. Sixteen sorts are available, and note that total payable and vehicle price sort separately because they are different numbers — cheapest total payable, dearest total payable, cheapest vehicle price, dearest vehicle price, newest and oldest listing, newest and oldest model year, lowest and highest mileage, smallest and largest engine, cars with or without inspection left first, and accident-free or repaired cars first. All five values we spot-checked reordered the first five rows. An unrecognised sort value breaks the call upstream with a 404, so it is validated here and comes back as INVALID_PARAM.

It means what is fitted, and we had to be careful about that. CarSensor renders the WHOLE equipment catalogue on every car's page and greys out the items that car does not have — on one measured car, 56 items listed and 17 actually fitted. Returning the list as-is would have claimed adaptive cruise control, lane-keep assist and a 360-degree camera on a car with none of them. So every row in equipment carries a fitted flag, and equipment_fitted is just the names that are really on the car, grouped by the site's own four categories (safety, comfort, interior, exterior). Filled on 12 of 12 records sampled.

The ones that actually decide a price in Japan. 修復歴 — whether the car has accident-repair history, which is the single biggest value signal in this market and which 489,368 of the 519,200 live cars declare clean. 車検 — how much of the compulsory roadworthiness certificate is left, because a buyer otherwise pays to renew it; you get the expiry month on the row and can filter on at least 6, 12 or 24 months remaining. 法定整備 — whether statutory servicing is included before handover. リサイクル料 — whether the recycling levy is already paid. 認定中古車 — manufacturer certification, on 107,861 cars. The era-dated registration year as the site prints it (2011(H23), Heisei 23). The kei-car class, which is 203,908 cars and a tax category, not a size. The 車台末尾番号, the chassis-number tail. And 福祉車両 welfare-vehicle and キャンピングカー camper classes with their own equipment codes. All of them filled 12 of 12 on the records we sampled.

CarSensor has no private sellers at all — every one of the 519,200 listings belongs to a registered dealer — so the dealer is always there. Every search row carries the dealer's trading name and full postal address (210 of 210 rows sampled, across seven different shelves). Every listing adds the dealer's numeric id and its resolvable dealer page. The dealer action then returns that dealer's profile: trading name, registered company name, postal address, opening hours, closing days, its own tagline, the badges CarSensor awards it, and the customer review scores published on the page with the sample size beside them. With include_stock you also get that dealer's entire advertised inventory in the same row shape as search — one extra call instead of thirty. One line against us: the only phone number the dealer page publishes is CarSensor's own free-dial number that forwards to the dealer, not the dealer's direct line, so it comes back as enquiry_phone with enquiry_phone_is_tracking_line true rather than as a telephone field you might trust.

Two actions exist for exactly that, and they return the live tables rather than a snapshot. makers returns all 119 maker codes with the maker's Japanese name, the country group CarSensor files it under and the number of cars it has listed right now — so you can size a brand before searching it (Toyota 118,750, Honda 74,460, Suzuki 72,148, Daihatsu 61,206, Nissan 55,726, Mercedes-Benz 9,598, Lexus 8,599). models returns one maker's model table with the same live count per model: 235 rows for Toyota led by Prius 9,196, 86 for Mazda led by CX-5 3,523. Several models in one search are a genuine OR — Prius at 9,196 plus Crown came back as 10,619. One honest note: CarSensor lists a maker's whole model history, including models nobody is selling today, and for those it prints no count — so live_listings comes back null rather than a made-up 0.

Things CarSensor does not publish, which we would rather name than have you discover. No sold prices and no price history — this is live asking-price inventory, so if you want a time series you have to collect it yourself. No depreciation curves or valuation estimates. No VIN: you get the chassis-number tail only, which is what the public page shows. No per-listing view count or posting timestamp. No facet counts alongside your results — the source prints one total per query and nothing else, so we do not invent a breakdown. No private sellers, because the site has none. No dealer geo-coordinates or aggregate rating, because the dealer page publishes neither. And no compact-car body type, because CarSensor's own backend rejects the option its own search panel offers.

Both are answers and both are typed properly. A listing that has sold or been withdrawn comes back as NOT_FOUND and is not retried: the source serves a real page saying 掲載終了, "listing ended", and that is a fact about the car rather than a problem with the request. A search that genuinely matches nothing returns ok with zero rows and total_results 0 — measured live, a city code under the wrong prefecture does exactly that. A malformed listing id is INVALID_PARAM, a missing one MISSING_PARAM, and a dealer id requested under the wrong prefecture is NOT_FOUND because the source keys the dealer page on the pair. 13 error cases were measured on two separate runs and all 13 landed on the right code both times.