Serbia's biggest car market, typed and filterable

75,467 live car ads on 2026-10-01 — 1,678 new, 73,789 used. You can reach all of them, but only if you ask for large pages, and this is the one number worth reading twice. The site advertises more pages than it will serve: the page number is capped at 750, and a request for page 751 or beyond comes back with page 1 again. At the default 25 ads per page that means 18,750 ads are reachable out of 75,467; at 200 ads per page, 750 pages covers 150,000, so the whole result set is yours. The API therefore accepts page_size 25, 50, 100 or 200, rejects page beyond 750 instead of handing you duplicate rows, and every response states pages (what the site claims), reachable_pages and reachable_results so you never have to find this out the hard way.

Checked two independent ways, not assumed. First, against the site's own published structured-data copy of the same adverts: 75 rows across three different searches, 74 prices matched exactly, 0 disagreed, and the currency matched on 74 of 74. The one row without a match had no price on either side — the seller published none. Second, a search row against the full ad for the same id on 8 cars: price, odometer and build year identical 8 of 8. Prices are read as the ad prints them and the currency is read from the ad too, never assumed: 900 rows sampled across the result set were all in euro, and if an ad ever prints something else the raw symbol comes back untranslated rather than relabelled. An ad with no price returns price null and price_on_request true — never a zero. That was 6 of 300 sampled rows.

All 35 of them were measured against the unfiltered total in the same run, with a 0.1 % tolerance because the live total drifts by a few dozen ads between calls. From 75,453 unfiltered: diesel 47,873, automatic 28,415, SUV 19,259, Euro 6 28,639, four-wheel drive 14,828, leather interior 13,955, dual-mass flywheel 14,211, imported from Germany 2,176, Audi 8,064, Audi A4 1,976, under EUR 3,000 11,291, 2020 or newer 12,352, under 50,000 km 4,627, 150 kW or more 8,589. Every one bit. Just as important, the filters the site accepts but quietly ignores are deliberately missing from this API: its equipment, safety, condition, selling-method, owner-status and region keys all left the total unchanged whichever way they were sent, and its horsepower box is ignored upstream, so power is filtered in kW only. A handle that silently returns the whole catalogue is worse than no handle.

The free-text description the seller wrote, the equipment list (the site tracks 97 distinct items; one sampled ad had 57 of them), the safety list (18 items tracked, 13 on that ad), every photo at full resolution rather than the one thumbnail, exterior colour, interior colour and upholstery material, drive layout, Euro class, whether the car is damaged, the registration validity and the site's own estimate of what registering it costs, the part-exchange stance, financing terms, reserved status, and the seller block with street, town, district, postal code, latitude and longitude, rating, active-ad count and the phone numbers the ad page shows. On the 6 ads sampled, 34 of 39 fields were filled on every one; description was on 5 of 6, and condition, owner legal status and selling method appear only when the seller filled them in, so they come back null instead of guessed.

brands returns all 99 brands with the exact spelling the search filter wants — and that matters, because the filter value is the site's label, 'Mercedes Benz' and 'Škoda' included, not a url slug. models is honest about a limit the site imposes: polovniautomobili publishes no model vocabulary anywhere on this surface, so models reads the brand's own live ads instead and returns the models that actually have cars for sale, each with the number of ads seen, plus brand_ads, ads_sampled and sample_coverage_pct so you can tell whether you got everything. Tesla returned Model 3, Model S, Model Y and Model X at 100 % coverage; Audi returned 31 models from three pages. A model with no live ad will not appear, and the response says so rather than pretending to be a catalogue.

Several, and they are the ones Serbian buyers filter on. Whether the car is on domestic plates, on foreign plates, or sold to be registered in the buyer's name — 51,577 of 75,453 were on domestic plates. Whether it has a dual-mass flywheel, which is a routine buying criterion there (14,211 ads). Whether it carries a conversion attest (3,026). Whether the vehicle is registered to the person selling it (37,165). The registration expiry date and the site's own estimate of the registration fee. The part-exchange stance, from 'no exchange' through 'for a cheaper car' to 'for a more expensive one'. Country of import, so you can separate German-sourced stock (2,176 ads) from the rest. And the gearbox is not just manual-or-automatic: the site distinguishes 4-, 5- and 6-speed manuals.

Measured across two consecutive full runs on 2026-10-01: 12 of 12 calls succeeded in each run, median 623 ms in one and 838 ms in the other, median 107 KB of upstream payload for a 25-ad page and 374 KB for a 200-ad page — one page is one upstream request either way, so asking for 200 ads costs one call, not eight. Field fill on 300 rows across five different searches: ad id, url, title, brand, model, year, currency, fuel, gearbox, body, doors, seats, power in kW and hp, town, photo and photo count all 100 %; displacement 99.3 %; odometer 99.0 %; price 98.0 %. Emission class is the honest exception on a search row — the site only attaches it to some rows, so it was filled on 12 of 300 (4.0 %). The full ad has it every time.

A removed or non-existent ad returns NOT_FOUND, not an empty success — Serbian ads come down as soon as the car sells, so this happens a lot and it is a terminal answer, not something to retry. A missing required parameter returns MISSING_PARAM, a bad enum value returns INVALID_PARAM before a request is spent, and an unknown action returns UNKNOWN_ACTION. Enum values are validated against the vocabulary the site itself publishes, which matters because of a trap worth knowing: the site's door filter only knows '3' (meaning 2/3 doors) and '5' (4/5 doors), and asking for 4 returns zero ads with no error at all. Here it returns INVALID_PARAM and tells you the two values that exist. filter_options returns all 24 vocabularies with their Serbian labels, and marks the groups the site publishes but does not actually filter on as filterable: false, with the reason.

Polovni Automobili API is a ReefAPI endpoint group for serbia's biggest car market as json — 75,467 live ads, 35 filters that actually bite. It returns live JSON through POST requests under /polovniautomobili/v1.

Yes. ReefAPI starts with 1,000 free credits, no card required. Polovni Automobili calls use the same shared credit balance as every other ReefAPI engine.

No login to Polovni Automobili is needed for the API response. You call ReefAPI with your x-api-key header, and the playground can run live examples before you create a production key.

The page example is captured from a live search call, and production requests fetch live data through ReefAPI rather than a static sample.

Polovni Automobili actions currently cost 2-3 credits per successful call. Failed or blocked calls are free. All APIs draw from one credit pool.

Yes. Connect ReefAPI once through MCP and your assistant can call polovniautomobili actions with the same key, credit pool and JSON envelope used by normal REST requests.