marketplace

Use this skill when the operator wants the full marketplace mount, routing metadata, or client setup. Trigger on "один MCP", "все источники", "marketplace_sources", "настроить marketplace", or "full m

SKILL.md

GitHub

Unified Marketplace Server

One server mounting every installed connector as a namespaced toolset: wb_*, ozon_*, yandex_*, detmir_*, avito_*, taobao_*, megamarket_*, lamoda_*, dns_*, citilink_*, aliexpress_*, cian_*, mpstats_* plus compare_prices / compare_sources and its own marketplace_sources and marketplace_health. Tool names keep their prefixes, so habits and configs carry over — but the operator wires a single marketplace entry instead of fourteen. The server exposes 41 tools with MPSTATS_MP_AUTH set and 39 without it: the two mpstats_* tools are not mounted until the paid token is configured, because without it every call would answer auth_missing.

When to use

  • "Where is X cheapest" — compare_prices fans out across all searchable sources
  • Client setup: one config entry, not fourteen
  • Health overview: the CLI's doctor runs every selfcheck at once
  • A source came back empty and you can't tell installed-but-quiet from never-loaded → marketplace_sources
  • Planning a fan-out, or a call just failed with a CDP/transport error → marketplace_health first: is Chrome up, is the host cooling down
  • A source wants a human at the browser → marketplace_health names the page

Tools

  • marketplace_sources() — no arguments; returns {mounted, skipped, mounted_count, skipped_count, capabilities, needs_cdp, server_version}. capabilities is static routing metadata per source: access tier, CDP/login requirements, currency, text-search support, and whether the connector is mounted. needs_cdp is the sorted subset of mounted whose reads can go through the operator's Chrome, so a dead browser is a reason to read marketplace_health before spending a call. skipped maps a source name to the reason it is unavailable: an import error (usually a missing dependency), a missing credential its tools need (MPStats without MPSTATS_MP_AUTH), or a MARKETPLACE_SOURCES deselection. Connectors are imported defensively, so a missing dep removes a marketplace instead of killing the server — but from the client an absent source looks identical to one that found nothing. Call this to tell the two apart: a name in skipped was never queried at all.
  • marketplace_health() — no arguments; read-only transport health without spending a navigation. Returns {cdp, budget, handoffs, unblock, server_version}. cdp is the scraping Chrome as it answers on the configured CDP address: reachable, host, port, version, page_targets, and a short reason/error when it is down. budget is the shared navigation budget: limits, in_flight, and per host {refusals, open, retry_after_s, opens} — an open host fails fast for retry_after_s more seconds. handoffs lists the retained challenge pages. unblock maps each mounted source with a page to the URL a human opens in the scraping Chrome. The tool never starts Chrome, opens or closes a tab, or touches a marketplace, and it never raises: a broken part is reported inside the answer. It does not prove a source is up — only a real call does.

Operator CLI

  • marketplace-mcp install [client] — print the exact mcpServers JSON to paste, or write it into the client's own config file with --write (the old file is kept as <file>.bak-<timestamp>). From a source checkout it prints the real filesystem path of that checkout, so there's no /path/to/ru-marketplace-mcp placeholder to hand-edit; installed as a wheel, it prints the console-script paths on PATH instead. The default block is one marketplace entry running the unified server; --per-source prints one entry per source. client must be one of claude, claude-code, cursor — an unknown name is rejected, not answered with a Claude block.
  • marketplace-mcp doctor [--status-file path] — per-source health plus a CDP session probe, with an optional machine-readable JSON snapshot for monitoring. Exit codes are meaningful for cron and CI: 0 everything healthy, 1 a parser drifted (the alarm — someone has to look), 2 nothing drifted but at least one source couldn't be judged (blocked, no CDP, wrong region). "Couldn't check anything" is a 2, never a 0.

Gotchas

  • Recovery-aware clients should read source_outcomes.requires_user_action before scheduling retries. Keep answered sources visible, show the affected marketplace and challenge_type, and retry that source after the browser interaction completes. A challenge is not an empty product search. See the compare-prices skill for the retry contract. handoff_expires_at confirms that the source retained its tab; repeat the same operation in the same MCP session after browser action and before that expiry. See docs/CDP_SETUP.md for opt-in configuration and supported sources.
  • A source whose optional deps are missing is simply absent from the set — marketplace_sources says which and why.
  • Set MARKETPLACE_SOURCES on the unified server to mount only a comma-separated subset (wildberries,ozon,compare); unset or blank keeps every source. The aliases wb, ym/yandex, detmir, and ali work too. Unknown names fail at startup instead of silently producing a partial server, and deselected sources are reported with a deselected reason.
  • compare_prices ranks on everyday ruble prices; Taobao (CNY) is reported in price_native but never ranked against rubles.

DSH activation

In DeepSeek Harness, the default profile exposes only compare_prices and compare_sources through the cheap compare mount. Per-marketplace tools, marketplace_sources and marketplace_health require RU_MARKETPLACE_MCP_FULL=1 and a profile restart; do not call them in the default mode.