marketplace · vladimir-human — Claude Code skill: setup and installation
A Claude Code skill for setting up one MCP server across marketplace connectors and routing requests. It sends price-comparison queries to compare-prices and skips single-source tasks.
MIT ★ 203 updated 2026-10-19
npx skills add Vladimir-Human/ru-marketplace-mcp --skill marketplace -a claude-codeWhat marketplace does
The skill explains how to connect one marketplace server instead of adding separate entries for each source. It describes marketplace_sources, which reports mounted and skipped sources, and marketplace_health, which returns browser transport and navigation-budget status. Price comparison requests go to compare-prices.
marketplace-mcp install [client] prints MCP settings or writes them to the client config with --write; marketplace-mcp doctor checks source health and the CDP session. Sources that rely on browser access need Chrome running with CDP configured. The CLI accepts only claude, claude-code, and cursor as client names.
MPStats tools mount only when the paid MPSTATS_MP_AUTH token is set. Without it, the two MPStats tools are skipped.
What it helps with
- connect one MCP server instead of separate source entries
- see which connectors mounted and which were skipped
- check the CDP session and browser transport status
- route a price comparison request to
compare-prices
Example prompts for the agent
- Set up one MCP server for all marketplace connectors in Claude Code.
- Show which sources mounted and which were skipped.
- Compare this item's price across the available sources.
How to install marketplace
1Skills CLI (Vercel)
npx skills add Vladimir-Human/ru-marketplace-mcp --skill marketplace -a claude-code2Manually: put the files in place
~/.claude/skills/marketplace/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/marketplace/
1Skills CLI (Vercel)
npx skills add Vladimir-Human/ru-marketplace-mcp --skill marketplace -a codex2Manually: put the files in place
.agents/skills/marketplace/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add Vladimir-Human/ru-marketplace-mcp --skill marketplace -a cursor2Manually: put the files in place
.cursor/skills/marketplace/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install Vladimir-Human/ru-marketplace-mcp skills/marketplace2Manually: put the files in place
.github/skills/marketplace/SKILL.md1Skills CLI (Vercel)
npx skills add Vladimir-Human/ru-marketplace-mcp --skill marketplace -a gemini-cli2Platform CLI
gemini skills install https://github.com/Vladimir-Human/ru-marketplace-mcp --path skills/marketplaceManually: put the files in place
.devin/skills/marketplace/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add Vladimir-Human/ru-marketplace-mcp --skill marketplace -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about marketplace
How do I install marketplace in Claude Code?
Run npx skills add Vladimir-Human/ru-marketplace-mcp --skill marketplace -a claude-code. The skill file is at ~/.claude/skills/marketplace/SKILL.md.
What does marketplace do?
It guides setup of a unified MCP server and explains how to inspect mounted sources and connection health. Price comparison requests are routed to compare-prices.
What does marketplace need to work?
Browser-backed sources need Chrome running with CDP configured. MPStats tools require the paid MPSTATS_MP_AUTH token.
Source text
SKILL.md · MIT
---
name: marketplace
description: 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 marketplace". Route price questions such as "где дешевле" and "compare prices" to compare-prices; skip single-source tasks.
---
# 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 mounts `compare-mcp`, which exposes
exactly 4 tools: `compare_prices`, `compare_sources`, `compare_verify_offer`
and `compare_browser_snapshot`.
The middle `decision-mcp` profile adds `decision_inspect` (5 tools). This
skill's own tools — `marketplace_sources` and `marketplace_health` — and every
per-marketplace tool need `RU_MARKETPLACE_MCP_FULL=1` and a profile restart;
do not call them in the default mode.
© Vladimir-Human · MIT · Vladimir-Human/ru-marketplace-mcp
Similar skills
All skills for Agent setup: memory, context, orchestration →marketplace by other authors comes first.