ru-marketplace-mcp

MCP servers for Russian and Chinese marketplaces. Prices, stock, ratings, reviews, and seller details from Wildberries, Ozon, Yandex Market, Detsky Mir, Avito, AliExpress, Taobao, MegaMarket, Lamoda, DNS, and Citilink. Plus real estate from Cian and price comparison across all product sources in a single call.

Read-only. No API keys, tokens, or registration needed — platforms with strict anti-bot protection are read through your own Chrome. One optional exception: MPStats takes a paid token (MPSTATS_MP_AUTH) — without it, everything else works as before.

English version below · Architecture · How to add a source · About anti-bot

An optional tab-persistence mode has been added for browser checks: CHROME_CHALLENGE_HANDOFF_S=120. After a check completes, repeating the same request within the same MCP session continues reading from that tab. Support and limitations are described in Chrome setup.

What's inside

Readable anonymously, without a browser: Wildberries, Yandex Market, Detsky Mir, and Lamoda product cards. Everything else needs your logged-in Chrome (CDP). Taobao and MegaMarket additionally require an active login to the platform itself — without it, Taobao hits the login wall, and MegaMarket returns an empty response. Avito also blocks by IP: from a datacenter address it's a hard refusal, from a Russian residential IP it works as long as you don't fire requests too fast. Requests to CDP sources are sent with spacing: a queue of consecutive requests without pauses will take them down (DNS and Taobao degraded that way in testing), so the connectors maintain a pause between calls on their own. The exact state from your session will be shown by marketplace-mcp doctor.

MPStats stands apart: it's the only paid source. Without MPSTATS_MP_AUTH the server still starts, but the tools respond with auth_missing — so it's optional and connected on demand, with no effect on the other thirteen servers.

A total of 39 tools across 14 servers on a shared mcp-core runtime. Plus a combined marketplace-mcp that mounts everything at once — one entry in the client config instead of fourteen. It adds its own tool marketplace_sources (which connectors came up and which went down and why), so it has 40 tools: 39 mounted plus this one.

Quick start

You need Python 3.12+ and uv.

git clone https://github.com/Vladimir-Human/ru-marketplace-mcp.git
cd ru-marketplace-mcp
uv sync --all-packages
uv run pytest -q -m "not live and not cdp"   # 1651 офлайн-тестов, сеть не нужна

Live endpoint check:

uv run python -c "
import asyncio
from wb_connector.server import wb_selfcheck
print(asyncio.run(wb_selfcheck()).status)   # ждём success
"

Connecting to an MCP client

Each server is a console command, so no paths are hard-coded in the config.

{
  "mcpServers": {
    "marketplace": {
      "command": "uv",
      "args": ["run", "--directory", "C:/путь/к/ru-marketplace-mcp", "marketplace-mcp"],
    },
  },
}

{
  "mcpServers": {
    "wildberries": {
      "command": "uv",
      "args": ["run", "--directory", "C:/путь/к/ru-marketplace-mcp", "wb-mcp"],
    },
    "ozon": {
      "command": "uv",
      "args": ["run", "--directory", "C:/путь/к/ru-marketplace-mcp", "ozon-mcp"],
    },
    "compare-prices": {
      "command": "uv",
      "args": ["run", "--directory", "C:/путь/к/ru-marketplace-mcp", "compare-mcp"],
    },
  },
}

{
  "mcpServers": {
    "marketplace": {
      "command": "uv",
      "args": ["run", "--directory", "C:/путь/к/ru-marketplace-mcp", "marketplace-mcp"],
      "env": {
        "MARKETPLACE_SOURCES": "wildberries,ozon,yandex_market,avito,aliexpress,dns,compare",
      },
    },
  },
}

claude mcp add wildberries -- uv run --directory /путь/к/ru-marketplace-mcp wb-mcp
claude mcp add yandex-market -- uv run --directory /путь/к/ru-marketplace-mcp yandex-mcp
claude mcp add detsky-mir -- uv run --directory /путь/к/ru-marketplace-mcp detmir-mcp
claude mcp add ozon -- uv run --directory /путь/к/ru-marketplace-mcp ozon-mcp
claude mcp add compare-prices -- uv run --directory /путь/к/ru-marketplace-mcp compare-mcp

{
  "mcpServers": {
    "compare-prices": {
      "command": "uv",
      "args": ["run", "--directory", "/путь/к/ru-marketplace-mcp", "compare-mcp"],
    },
  },
}

dsh plugin --profile web add github:Vladimir-Human/ru-marketplace-mcp#path:/dsh

After connecting, restart the client and run marketplace-mcp doctor. It launches each connector's canary and answers success, drift_detected, or inconclusive.

Tools

The *_selfcheck canaries are intentionally not in this list: they are not published via MCP because operator diagnostics would cost the model ~7.5k tokens in every request. marketplace-mcp doctor runs them — all at once, from the command line.

Wildberries — wb_*

wb_seller answers the question a product card hides: who is actually selling? It returns the registered legal entity and tax numbers. This is how you tell an official brand store from a reseller with a similar name.

wb_questions fills another gap. Reviews tell you what it's like to own the product; questions clarify what the product actually is — "10 or 16 amps", "cable included?". The seller's answer is often the only public statement about it. The pool is shared across all product variants, key is imt_id from wb_root_info.

wb_category_products completes the chain with wb_categories: the latter gives you shard and query, the former — products by them. Item format matches wb_search, so category browsing and text search are directly comparable. Some large WB sections are marked with the blackhole shard — they have no results of their own, and the tool honestly says so instead of returning an empty list.

Yandex Market — yandex_*

Two prices, always. price_rub is what any buyer pays. price_with_plus requires a Yandex Plus subscription and is usually 25–30% lower. Yandex's UI shows the second one in large font, so calling it without a caveat means promising a price that a person without a subscription won't get.

A search result row is an offer from the listing, not the default offer of the card. A single product_id covers a product family, and the listing may show one representative while the card for the same id shows another; compare search rows to cards by sku_id, not by URL. price_old_rub in search rows is the strikethrough base price, discount context; it cannot be called the price.

rating_stars gives a distribution like {1: 10, 2: 3, 3: 10, 4: 19, 5: 502}. From it you can tell whether an average of 4.8 is honest or if a pile of 1-star reviews is hiding behind it.

Detsky Mir — detmir_*

Region is specified per call. Prices and especially in-store availability depend heavily on the city: the same product was in 152 stores in Moscow, 37 in Saint Petersburg, and 2 in Khabarovsk. The region parameter overrides DETMIR_REGION, so you can compare cities within one session.

There is no text search here, and that's intentional. Detsky Mir's API silently ignores any text filters and returns the entire catalog of 300k items, while the site's search route returns a 404 with a promo carousel. A search tool would return confidently wrong products, so navigation goes through categories. Details in docs/ANTI_BOT.md.

Ozon — ozon_*

Ozon rejects datacenter traffic, so the connector is two-tier. First, TLS impersonation. If Cloudflare issues a challenge, the request is executed inside your logged-in Chrome via DevTools Protocol. Nothing is stored: you log in yourself, in a browser you control. Setup is described in docs/CDP_SETUP.md.

From a Russian residential IP the first tier usually works and no browser is needed.

Ozon reviews are shared across the whole product family card, and pool neighbors — often a different product from a different brand. On the Huter 1500 W oil radiator card (SKU 5264146973, rating 4.8 from 356 reviews), among 100 pulled reviews there was not a single one about the actual Huter: 38 about Resanta 2000 W, 34 about Resanta 1500 W, 5 about Eurolux, and so on — 12 products in the pool total. That's why every review carries item_id — the SKU of the product it was written about, and the response additionally returns requested_item_id, own_reviews (how many reviews are actually about this SKU), and pool_variants (SKU → название for every product in the pool). rating_score and distribution are computed over the pool, not the product: before drawing a conclusion, reviews must be filtered by item_id, and when own_reviews: 0 — honestly state that the product has no own reviews.

Avito — avito_*

Avito is listings, not a catalog: there's no product review pool, and seller reputation is the trust signal. A free/exchange listing comes back with price_rub: null — never 0, so it doesn't end up as "cheapest" in a comparison. From a datacenter IP Avito answers with a 403 firewall, so the connector is two-tier: TLS impersonation, then your Chrome (as with Ozon).

Taobao — taobao_*

Taobao search — client-side React app with a signed mtop API: every request requires a sign computed from the cookie token, so there's no anonymous path. All reads happen inside your Chrome, where the site signs requests itself. Prices are in CNY and are not converted: a hardcoded rate would silently go stale, so compare with ruble-based sources explicitly.

MegaMarket, Lamoda, DNS, Citilink

These four are read via your Chrome (CDP). MegaMarket (megamarket_*) — mobile JSON API behind ServicePipe, and a single passed challenge isn't enough: the API gives an anonymous session an empty list; an active MegaMarket login is required. DNS (dns_*) and Citilink (citilink_*) — rendered DOM behind Qrator; all three have no anonymous path at all. Lamoda (lamoda_*) is half-and-half: cards are fetched anonymously via GraphQL, while search goes through Chrome. Chrome with CDP (scripts/start_chrome_cdp.sh) is needed for all of them except Lamoda cards.

In total, eight sources go through CDP — these plus Taobao, AliExpress, Ozon, and Avito, where Chrome is just a fallback tier: their tier 1 usually responds, and the browser kicks in when the anonymous tier hits a challenge. marketplace-mcp doctor from your browser will tell you which endpoints are confirmed.

AliExpress — aliexpress_*

Read via your Chrome (CDP): x5sec serves a captcha to anonymous clients, so the connector lands on the search page (it's not challenged) and opens the card in a new tab from there. Prices are in rubles and participate in compare_prices. A card with a name but no price is a known state: under load x5sec stops returning the price module, and the connector reports price_missing rather than inventing a number. The "N ₽ with coupon" price is not published in price_rub: that field carries the regular price, and the connector honestly warns about the coupon separately. Review texts are not returned: only rating and order count. As with the other CDP sources, a green aliexpress_selfcheck proves the transport responded — not that the price is correct.

Cian — cian_*

Недвижимость, а не товары: квартиры, комнаты, дома и коммерция на продажу, в долгосрочную аренду и посуточно (deal="daily"; коммерции посуточно у Циана нет, такой запрос отклоняется). Длительная и суточная аренда — два разных рынка, в одной выдаче не смешиваются: у суточных цена за ночь, price_period и lease_term пустые, поэтому у каждой строки есть price_unit — total, month или day. Сравнивать цены между единицами нельзя. Поиск только по фильтрам — текстового поиска у Циана нет. Регион задаётся id Циана: 1 Москва, 2 Санкт-Петербург, 4593 Московская область, 4588 Ленинградская область (все четыре проверены живьём); остальным регионам нужен их id. Читается через ваш Chrome (CDP): WAF Циана режет голый HTTP по IP, а из сессии браузера отвечает собственный JSON-API сайта, так что HTML не парсится. Цена «не указана» приходит как null, не 0. Агентской страницы как инструмента нет: она не отдаёт структурированных данных, агент приходит внутри карточки. В compare_prices источник не участвует.

Сравнение цен — compare_*

compare_prices("кроссовки мужские")

  wildberries      712 ₽   Кроссовки изи дышащие спортивные
  wildberries      814 ₽   Зимние кроссовки теплые с мехом
  yandex_market   2499 ₽   Кеды A-LOW
  yandex_market   3480 ₽   Кеды

  дешевле всего: wildberries 712 ₽, разброс 5858 ₽, complete: true

Маркетплейсы опрашиваются параллельно, и каждый отчитывается сам за себя. Если один заблокирован, сравнение не рушится: complete: false вместе с source_outcomes покажет, что именно вы видите. Подписочные цены в ранжировании не участвуют. Совпадающие предложения по паре (источник, id товара) схлопываются, так что один и тот же товар не занимает два места в ранжировании.

У каждого предложения есть currency (строчный ISO-код, по умолчанию rub) и price_native — цена в этой валюте, как её показывает маркетплейс. Для российских источников она совпадает с price_rub; у Taobao в ней лежит цена в юанях, которую price_rub намеренно оставляет пустой. Раньше юаневую цену забирали и молча выбрасывали, и строка Taobao приходила с пустой ценой без намёка, что цена вообще есть. Теперь юань виден, но в рублёвом ранжировании по-прежнему не участвует: в warnings появляется foreign_currency: … с числом исключённых предложений и причиной. Конвертировать здесь значило бы зашить курс, который молча устареет, — пересчёт за вами.

MPStats — mpstats_*

Аналитика продаж и остатков по SKU Ozon и Wildberries через плагин MPStats. В отличие от всех остальных коннекторов, этот опционален и требует платный аккаунт MPStats: авторизация — одна cookie mp_auth (JWT из залогиненной сессии плагина на mpstats.io), задаётся переменной MPSTATS_MP_AUTH. Без неё инструменты возвращают auth_missing, а сервер запускается как обычно — ни на что другое это не влияет.

place — ozon или wildberries. Графики длиной 30, от старых к новым: последняя ненулевая ячейка — текущая цена или остаток. Цена и остаток при сплошь нулевом графике ведут себя намеренно по-разному: цена становится None (ложный 0 выиграл бы любое сравнение «где дешевле»), а остаток — 0, потому что «нулевой остаток» это осмысленное показание, а не отсутствие данных. Пустой график даёт None в обоих случаях. Ноль в отдельной ячейке — «нет данных за тот день», а не «значение было нулевым», поэтому сумму за окно считайте по графику. Отсутствие токена и транспортные сбои selfcheck отчитывает как inconclusive, не drift: гоняться за дрейфом схемы, которого не было, не нужно. Токен — секрет платного аккаунта с квотой: не логируйте и не коммитьте его.

Навыки для агента

У каждого коннектора — свой навык в skills/: пятнадцать штук, по одному на источник плюс общий marketplace. Навык это не пересказ README: он объясняет агенту, когда за этот источник вообще браться, чего у источника нет, и каким его ответам нельзя верить без второго взгляда.

mcp-core — общий рантайм под остальными серверами. Своего навыка у него нет.

Соответствие проверяется тестом (packages/marketplace-connector/tests/test_skills_parity.py): новый коннектор без навыка роняет прогон, как и навык, который называет несуществующий инструмент или забыл существующий. До этого теста навык DNS почти год советовал формат ссылки /product/<24-hex>/ — тот самый шаблон, который чинили как баг.

Скиллы едут в Docker-образ (/app/skills/), но в колёсах их нет: skills/ лежит в корне репозитория. Ставите с PyPI — возьмите навыки из репозитория отдельно.

Настройка

Все параметры задаются переменными окружения с префиксом коннектора. Все необязательные.

CHROME_CDP_HOST указывает, куда дозвониться CDP-клиенту (по умолчанию 127.0.0.1). Из контейнера ставьте chrome (сайдкар) или host.docker.internal — это открывает tier-2 источники (Ozon, Авито, Taobao, Мегамаркет, Lamoda, DNS, Ситилинк) в Docker без host networking. Подробности в docs/DEPLOYMENT.md.

*_CACHE_TTL=0 выключает кэш. *_PROXY перекрывает стандартные HTTPS_PROXY и ALL_PROXY — свой префикс есть у семи коннекторов: WB_, YANDEX_, DETMIR_, OZON_, AVITO_, LAMODA_ и MPSTATS_. У Taobao своего нет намеренно: поиск там подписан и ходит через собственный клиент. У Мегамаркета, DNS и Ситилинка тоже нет: их трафик идёт через ваш Chrome, а его egress — дело настроек браузера. Кэшируются только удачные ответы: запомнить сбой значило бы растянуть секундную помеху на весь TTL.

У Ozon прокси применяется к первому уровню. Второй идёт через ваш собственный Chrome, и его трафик — дело настроек этого браузера.

Секрет один, и тот опциональный. Всем серверам, кроме MPStats, ничего не нужно: нечего настраивать, нечему утечь. У MPStats есть MPSTATS_MP_AUTH — JWT платного аккаунта, и потому его место только в env клиентской записи: в коде и коммитах его нет и быть не должно.

Разработка

uv sync --all-packages
uv run pytest -q -m "not live and not cdp"    # 1651 офлайн-тестов
uv run pytest -q -m "not live"                # то, что гоняет CI
uv run pytest -q -m "not live" --cov          # покрытие, порог 70% в CI
uv run ruff check . && uv run ruff format --check .
uv run mypy                                   # что проверять — в [tool.mypy] files
uv run mypy --platform win32                  # ловит ошибки, видимые только на Windows
uv run python scripts/check_no_print.py       # запись в stdout ломает JSON-RPC
uv run python scripts/check_versions.py       # одна версия во всех 77 местах

Часть тестов прогоняет настоящий JS-экстрактор коннектора по снятой разметке и проверяет результат против цен, которые в тот момент были на странице. Для этого нужен Node с jsdom:

npm install jsdom      # либо NODE_PATH на уже установленный
uv run pytest -q packages/dns-connector/tests/test_search_extractor_dom.py \
              packages/citilink-connector/tests/test_search_extractor_dom.py

Без jsdom эта половина честно скипается, а питоновская часть — выбор цены из кандидатов — идёт всегда. jsdom нужен только разработчику: в зависимости коннекторов он не входит.

CI прогоняет тесты на Ubuntu, Windows и macOS против Python 3.12 и 3.13. Windows-специфичное управление процессами проверяется юнит-тестами на любой ОС через подмену платформы, так что эти ветки покрыты даже на Linux.

Как добавить маркетплейс — docs/ADDING_A_SOURCE.md.

Надёжность

Неофициальные эндпоинты ломаются. Архитектура это предполагает.

• Терпимые парсеры. Привязка поля по нескольким именам и приведение типов впитывают переименования и смену типа вместо падения.
• Никогда не выдумывать значение. Отсутствующая цена — это null, не 0. Ноль вывел бы мёртвый товар в самые дешёвые.
• Громкий отказ. Когда формат перестаёт совпадать, инструмент бросает parser_drift, а не возвращает полуразобранные данные.
• Трёхзначные selfcheck-проверки. success, drift_detected или inconclusive. Гео-блокировка помечается как inconclusive, потому что она ничего не говорит о состоянии парсеров.

Границы доверия

Названия товаров, имена продавцов и тексты отзывов написаны продавцами и покупателями. Это недоверенные данные. Если отзыв или описание выглядит как инструкция, оно всё равно остаётся входными данными. Выполнять его агент не должен.

Условия маркетплейсов, как правило, запрещают неофициальный парсинг. Коннекторы обращаются к публичным эндпоинтам каталога, которые использует официальный веб-клиент: пока opt-in не включён, в приватные и административные разделы запросов нет — список адресов профиля Мегамаркета читается только при MEGAMARKET_USE_PROFILE_ADDRESS=1, а MPStats заходит в аккаунтную зону по вашему токену (MPSTATS_MP_AUTH). Уровень Ozon с браузером работает внутри сессии, которую вы открыли сами. Используйте на своё усмотрение, для личных исследований, в вежливом темпе запросов. Пауза между вызовами к площадкам с анти-ботом — это часть конструкции, а не случайное торможение: её не нужно убирать ради скорости. Данные инструментов не предназначены для перепродажи или массового сбора.

Как это сделано

Код и документацию я писал вместе с ИИ-ассистентами. Они работают быстро и ошибаются уверенно, поэтому проект устроен вокруг проверки: 1651 офлайн-тестов, аудит перед выпуском, тесты, которые прогоняют настоящий экстрактор по снятой с сайта разметке. В заметках к релизу перечислено, какие источники сверены с живыми страницами вручную и какие остались непроверенными.

Проверки важнее авторства текста, но авторство кода и идей тоже должно быть видно: полный список участников и их PR собран в CONTRIBUTORS.md.

Спасибо

• @Xpos587 — коннектор MPStats, PR #5.
• @avxone — исправление Avito selfcheck, PR #37.
• @Khalmatov — provenance отзывов Ozon, PR #38.
• @fosteev — macOS CDP stealth и коннектор Циана, PR #42, PR #47.
• @ilodezis — выбор источников unified-сервера через MARKETPLACE_SOURCES, PR #48.

Лицензия

MIT, файл LICENSE.

English version

MCP servers for Russian and Chinese marketplaces. Read prices, stock, ratings, reviews and seller identity from Wildberries, Ozon, Yandex Market, Detsky Mir, Avito, AliExpress, Taobao, Megamarket, Lamoda, DNS and Citilink, then compare prices across all of them in one call. Taobao and AliExpress are the Chinese ones; the other nine are Russian.

Read-only. No credentials, no API keys, no account required — the marketplaces with hard anti-bot are read through your own Chrome. One optional exception: MPStats takes a paid account token (MPSTATS_MP_AUTH) if you want its analytics; without it every other server is unaffected.

What you get

Anonymous, no browser: Wildberries, Yandex Market, Detsky Mir and Lamoda cards. The rest need your logged-in Chrome (CDP). Taobao and Megamarket additionally need you signed into the marketplace itself — without it Taobao hits a login wall and Megamarket returns an empty result. Avito also blocks by IP: from a datacenter address it is a flat refusal, from a Russian residential one it works as long as you do not burst requests. Requests to the CDP sources are paced apart — a run of back-to-back calls degrades them (DNS and Taobao both dropped that way in testing), so the connectors hold a gap between calls themselves. Run marketplace-mcp doctor from your own session for the current state.

MPStats stands apart as the only paid source: without MPSTATS_MP_AUTH the server boots but its tools answer auth_missing. It is therefore optional — plug it in if you have an account; the other thirteen servers never notice.

39 tools across 14 stdio MCP servers, sharing one runtime (mcp-core), plus the unified marketplace-mcp that mounts them all under one client entry. It adds its own marketplace_sources tool — which connectors mounted, and which dropped out and why — so it exposes 40 tools: the 39 mounted plus that one. stdio is the default; HTTP transport is opt-in for remote deployment — see docs/DEPLOYMENT.md.

Quickstart

Requires Python 3.12+ and uv.

git clone https://github.com/Vladimir-Human/ru-marketplace-mcp.git
cd ru-marketplace-mcp
uv sync --all-packages
uv run pytest -q -m "not live and not cdp"    # 1651 offline tests, no network needed

Client configuration mirrors the Russian section above. Each server is a console script (wb-mcp, ozon-mcp, yandex-mcp, detmir-mcp, aliexpress-mcp, cian-mcp, compare-mcp) launched through uv run --directory /path/to/repo <script>. The optional mpstats-mcp runs the same way with MPSTATS_MP_AUTH in the entry's env (paid MPStats account; without it the tools return auth_missing). marketplace-mcp install [claude|claude-code|cursor|dsh] prints the block with your checkout's real path filled in — no placeholder to hand-edit — or the console-script paths on PATH when installed as a wheel; an unknown client name is rejected. The dsh target prints a cordis.patch.yml row instead of mcpServers JSON — see dsh/README.md.

DeepSeek Harness (dsh) installs as a plugin bundle rather than an mcpServers entry, from the dsh/ subdirectory (pnpm must be on PATH):

dsh plugin --profile web add github:Vladimir-Human/ru-marketplace-mcp#path:/dsh

That gives you 15 skills immediately and no MCP tools: both MCP rows stay disabled until RU_MARKETPLACE_MCP_DIR points at a clone. A mounted server is paid on every request — ~0.9k tokens for the recommended price-comparison mode, ~13.6k for the full set — so opting in is left to you. dsh/README.md covers enabling it and the full mode.

After connecting, run marketplace-mcp doctor. It runs every connector's canary and reports success, drift_detected, or inconclusive for each.

The tools

The *_selfcheck canaries are deliberately absent from these tables: they are not published over MCP, because operator diagnostics would cost the model ~7.5k tokens on every request. marketplace-mcp doctor runs them all from the command line.

Wildberries — wb_*

wb_seller answers the question a listing hides: who actually ships this? It returns the registered legal entity and tax ids, which is how you distinguish an official brand store from a reseller trading under a lookalike name.

wb_questions covers a different gap. Reviews describe what owning the product is like; questions clarify what it actually is — "10A or 16A?", "is the cable included?" — and the seller's reply is often the only public statement of that fact. One pool per imt_id, shared across every variant.

wb_category_products closes the loop wb_categories opens: that tool hands back WB's shard and query, and this one fetches the products behind them. Items use the same shape as wb_search, so a category walk and a text search are directly comparable. Several of WB's largest sections carry the shard blackhole and have no feed at all; the tool says so instead of returning an empty list.

Yandex Market — yandex_*

Two prices, always. price_rub is what anyone pays. price_with_plus needs a paid Yandex Plus subscription and runs 25–30% lower. Yandex leads with the subscriber price, so quoting it uncritically misstates the real cost.

A search row is the SERP's offer, not the card's default offer. One product_id covers a product family, and search may show one member while the card for the same id defaults to another — reconcile rows with cards by sku_id, not by URL. price_old_rub on search rows is the struck-through reference price, never a price to quote.

rating_stars gives the distribution, for example {1: 10, 2: 3, 3: 10, 4: 19, 5: 502}. That reveals whether a 4.8 average is earned or hides a cluster of complaints.

Detsky Mir — detmir_*

Region is per call. Prices and especially offline availability swing by city — one item sat in 152 Moscow stores, 37 in St Petersburg, 2 in Khabarovsk. The region parameter overrides DETMIR_REGION, so one session can compare cities.

There is no text search, deliberately. Detsky Mir's API silently ignores every text filter and returns its entire 300k-item catalog; the website's search route answers 404 and renders a promo carousel. A search tool would return confidently wrong products, so discovery goes through categories instead. See docs/ANTI_BOT.md.

Ozon — ozon_*

Ozon rejects datacenter traffic, so this connector is two-tier: TLS impersonation first, then a fetch inside your own logged-in Chrome over the DevTools Protocol when Cloudflare challenges. Nothing is stored; you log in yourself, in a browser you control. Setup: docs/CDP_SETUP.md.

From a Russian residential IP the first tier usually works and no browser is needed.

Ozon pools reviews per card family, and the neighbours are often a different product from a different brand. A live card for a 1500 W Huter oil heater (SKU 5264146973, rated 4.8 across 356 reviews) returned 100 reviews of which zero were about the Huter: 38 about a 2000 W Resanta, 34 about a 1500 W Resanta, 5 about a Eurolux — 12 products in that pool. So every review carries item_id, the SKU it is actually about, and the response adds requested_item_id, own_reviews (how many returned reviews really are about that SKU) and pool_variants (SKU → name for the whole pool). rating_score and distribution are pool-wide, not per product: filter by item_id before concluding anything, and when own_reviews is 0, say plainly that the product has no reviews of its own.

AliExpress — aliexpress_*

Read through your Chrome (CDP): x5sec challenges anonymous clients, so the connector lands on a search page (never challenged) and opens the card in a new tab from it. Prices are rubles and rank in compare_prices. A card with a title but no price is a known state — under load x5sec stops serving the price module and the connector reports price_missing rather than inventing a number. A "with coupon" price never lands in price_rub: the regular price does, and the coupon is reported as a warning. Review texts are not exposed; rating and order counts are. As with every CDP source, a green aliexpress_selfcheck proves the transport answered — not that a given price is right.

Cian — cian_*

Real estate, not goods: flats, rooms, houses and commercial property for sale, long-term rent or daily rent (deal="daily"; Cian has no daily commercial market and that combination is refused). Long-term and daily are separate markets that never share a result page: a daily price is per night and Cian leaves price_period and lease_term null there, so every row carries price_unit — total, month or day. Never compare prices across units. Search is by filters only — Cian has no text search. The region is a Cian id: 1 Moscow, 2 St. Petersburg, 4593 Moscow oblast, 4588 Leningrad oblast (all four verified live); other regions need their own id. Read through your Chrome (CDP): Cian's WAF blocks plain HTTP by IP, while inside the browser session the site's own JSON API answers, so no HTML is parsed. A price Cian does not state arrives as null, never 0. There is no agent tool: the agent page exposes no structured data, and the publisher ships inside the card. The source takes no part in compare_prices.

Avito — avito_*

Avito is classifieds, not a catalog: there is no per-product review pool, the seller's reputation IS the trust signal. A free/swap listing arrives with price_rub: null — never 0, so it cannot win "cheapest". From a datacenter IP Avito answers a 403 firewall, hence the two-tier transport: TLS impersonation first, then your Chrome, exactly like Ozon.

Taobao — taobao_*

Taobao search is a signed-mtop React app: every request needs a sign derived from a cookie token, so there is no anonymous path. All reads run inside your Chrome, where the site signs its own requests. Prices stay in yuan (CNY) and are never converted — a baked-in rate quietly goes stale, so compare ruble and yuan listings explicitly.

Megamarket, Lamoda, DNS, Citilink

These four read through your Chrome (CDP). Megamarket (megamarket_*) goes through the mobile JSON API behind ServicePipe and needs an active login — an anonymous session reads empty. DNS (dns_*) and Citilink (citilink_*) render DOM behind Qrator with no anonymous path at all. Lamoda (lamoda_*) is split: cards over anonymous GraphQL, search through Chrome. All of them need Chrome with CDP (scripts/start_chrome_cdp.sh), except Lamoda cards. Eight sources run through CDP in total: these plus Taobao, AliExpress, Ozon and Avito, where Chrome is only the fallback tier when the anonymous one is challenged.

Cross-marketplace — compare_*

compare_prices("кроссовки мужские")

  wildberries      712 RUB   Кроссовки изи дышащие спортивные
  wildberries      814 RUB   Зимние кроссовки теплые с мехом
  yandex_market   2499 RUB   Кеды A-LOW
  yandex_market   3480 RUB   Кеды

  cheapest: wildberries 712 RUB, spread 5858 RUB, complete: true

Sources are queried concurrently and each reports its own outcome. One marketplace being blocked never sinks the comparison: complete: false plus source_outcomes tells you exactly what you are looking at. Subscription prices never win the ranking. Offers matching on (source, product id) are collapsed, so one listing can no longer take two ranking slots.

Every offer carries currency (lowercase ISO code, default rub) and price_native, the price in that currency as the marketplace quotes it. For Russian sources it mirrors price_rub; for Taobao it holds the yuan price that price_rub deliberately leaves null. That yuan price used to be fetched and silently thrown away, so a Taobao row showed a blank price with no sign a real one existed. Now the yuan is reported but still never ranked against roubles: a foreign_currency: … warning lists how many offers were excluded and why. Converting here would bake in an exchange rate that goes stale silently, so the caller converts if they want to.

MPStats — mpstats_*

Sales and stock analytics per Ozon or Wildberries SKU via the MPStats browser plugin. Unlike every other connector, this one is optional and needs a paid MPStats account: auth is a single mp_auth cookie (JWT from a logged-in plugin session at mpstats.io), set via the MPSTATS_MP_AUTH env var. Without it the tools return auth_missing while the server boots normally — nothing else is affected.

place is ozon or wildberries. Graphs are length 30, oldest first: the last non-zero cell is the current price or stock. The two differ on purpose when the whole graph is zero: price becomes None (a false 0 would win any "cheapest" comparison), while stock becomes 0, because "none in stock" is a real reading rather than an absence of data. An empty graph yields None for both. A zero cell means "no data for that day", not "the value was zero", so sum the graph for a window total. A missing token or a transport failure reports as inconclusive, not drift — no chasing a schema drift that never happened. The token is a secret on a paid, quota-billed account: never log or commit it.

Agent skills

Every connector ships its own skill under skills/ — fifteen of them — one per source plus a shared marketplace overview. A skill is not a restatement of this README: it tells the agent when to reach for that source at all, what the source does not have, and which of its answers should not be trusted without a second look.

mcp-core is the shared runtime rather than a server, so it has no skill.

The mapping is enforced by a test (packages/marketplace-connector/tests/test_skills_parity.py): a new connector without a skill fails the run, and so does a skill that names a tool which does not exist — or omits one that does. Before that test existed, the DNS skill spent months telling operators to pass /product/<24-hex>/, the exact pattern a fix had already removed.

Skills are copied into the Docker image (/app/skills/), but they are not in the wheels: skills/ lives at the repository root rather than inside the packages. Installing from PyPI means fetching the skills from the repo separately.

Configuration

Every setting is an environment variable with a per-connector prefix. All optional.

*_CACHE_TTL=0 disables caching. *_PROXY overrides the standard HTTPS_PROXY/ALL_PROXY — seven connectors carry one: WB_, YANDEX_, DETMIR_, OZON_, AVITO_, LAMODA_ and MPSTATS_. Taobao has none by design and Megamarket, DNS and Citilink none either: their traffic goes through your own Chrome, whose egress is that browser's configuration. Only successful reads are cached: remembering a failure would stretch a one-second blip across the whole TTL window.

Ozon's proxy applies to tier 1. Tier 2 runs inside your own Chrome, whose egress is that browser's configuration, not ours.

Containers. CHROME_CDP_HOST points the CDP client at Chrome (default 127.0.0.1; use chrome or host.docker.internal inside Docker). That single variable is what opens the tier-2 sources — Ozon, Avito, Taobao, Megamarket, Lamoda, DNS, Citilink and AliExpress — from a container without host networking. See docs/DEPLOYMENT.md.

One secret, and it is optional. Every server except MPStats needs nothing: nothing to configure, nothing to leak. MPStats alone has MPSTATS_MP_AUTH, the JWT of a paid account — it belongs only in the client entry's env, never in code or commits.

Only the sources you use: MARKETPLACE_SOURCES. The unified server mounts every source, and their tool schemas are sent to the client on every request. Set MARKETPLACE_SOURCES in the marketplace-mcp entry's env to a comma-separated list to mount only those, e.g. wildberries,ozon,yandex_market,avito,aliexpress,dns,compare. Names are canonical (wildberries, ozon, yandex_market, detsky_mir, avito, taobao, megamarket, lamoda, dns, citilink, aliexpress, cian, compare, mpstats); the aliases wb, ym/yandex, detmir and ali work too. An unknown name is rejected at startup with the supported-source list, so a typo cannot silently produce a partial server. Unset or blank mounts everything, as before. Deselected sources show up in marketplace_sources under skipped, marked as deselected rather than failed to import, and compare_prices queries the same subset.

Development

uv sync --all-packages
uv run pytest -q -m "not live and not cdp"    # 1651 offline tests
uv run pytest -q -m "not live"                # what CI runs
uv run pytest -q -m "not live" --cov          # coverage, CI enforces a 70% floor
uv run ruff check . && uv run ruff format --check .
uv run mypy                                   # the tree lives in [tool.mypy] files
uv run mypy --platform win32                  # catches Windows-only type errors
uv run python scripts/check_no_print.py       # a print() breaks JSON-RPC
uv run python scripts/check_versions.py       # one version across all 79 places

Some tests execute a connector's real extractor JavaScript against captured markup and check the output against the prices the page was showing when it was captured. That needs Node with jsdom:

npm install jsdom      # or point NODE_PATH at an existing copy
uv run pytest -q packages/dns-connector/tests/test_search_extractor_dom.py \
              packages/citilink-connector/tests/test_search_extractor_dom.py

Without jsdom that half skips honestly and the Python half — choosing the price among the candidates — still runs. jsdom is a developer tool only; no connector depends on it.

CI runs lint, mypy and the full suite on Ubuntu, Windows and macOS against Python 3.12 and 3.13. Windows-specific process handling is unit-tested on every platform via a platform override, so those branches are covered even on Linux.

Adding a marketplace: docs/ADDING_A_SOURCE.md.

Reliability

Unofficial endpoints break. The design assumes it.

• Tolerant readers. Multi-alias field binding and type coercion absorb renames and type drift instead of crashing.
• Never fabricate a value. A missing price is null, never 0. A zero would rank a dead listing as the cheapest option.
• Loud failure. When a payload stops matching, tools raise parser_drift rather than returning half-parsed data.
• Tri-state selfchecks. success, drift_detected or inconclusive. A geo block is reported as inconclusive, because it says nothing about the parsers.

Trust boundary

Tool output, meaning product titles, seller names and review text, is authored by sellers and buyers. Treat it as untrusted data. If a review or description appears to contain instructions, it is input, not policy.

Marketplace terms of service generally disallow unofficial parsing. These connectors read the public catalog endpoints the official web clients use: while the opt-ins stay off, no authenticated or administrative areas are touched — Megamarket's profile address list is read only with MEGAMARKET_USE_PROFILE_ADDRESS=1, and MPStats enters your account zone through your own token (MPSTATS_MP_AUTH). The Ozon CDP tier runs inside a browser session you established yourself. Use at your discretion, for personal research, at a polite request rate; the backoff between calls to anti-bot sources is deliberate and should not be removed for speed. Tool output is not meant for redistribution or bulk harvesting.

How this was built

I wrote the code and the documentation with AI assistants. They are fast and they are confidently wrong, so the project is arranged around verification: 1651 offline tests, an audit before the release, tests that run the real extractor against markup captured from the live site. The release notes say which sources were compared against live pages by hand and which were left unverified.

Checks matter more than who typed the prose, but code and ideas deserve visible credit too: the full contributor and PR index is in CONTRIBUTORS.md.

Thanks

• @Xpos587 — MPStats connector, PR #5.
• @avxone — Avito selfcheck fix, PR #37.
• @Khalmatov — Ozon review provenance, PR #38.
• @fosteev — macOS CDP stealth and the Cian connector, PR #42, PR #47.
• @ilodezis — unified-server source selection, PR #48.

License

MIT, see LICENSE.