marketua
Search Ukrainian marketplaces (OLX, Hotline, Prom; Rozetka via Bright Data MCP) from a plain-language prompt and get structured offers with prices and links.
README
marketua
An MCP server that lets AI agents (Claude Desktop, Hermes Agent, Cursor, OpenCode, …) search Ukrainian marketplaces — OLX, Hotline.ua, Prom.ua and Rozetka — from a plain-language prompt and get back structured offers with links.
Status: alpha. OLX, Hotline and Prom work out of the box (no tokens); Rozetka is covered through the optional Bright Data MCP server (free tier: 5 000 requests/month). Verified with MCP Inspector and a real agent (OpenCode), including a 4-marketplace price comparison from a single prompt.
Idea
Ask in natural language:
"SSD 120–560 GB, 2.5", used, up to 1500 UAH"
and get a structured list of matching offers — title, price, location, link — instead of a wall of search results.
Why not just use web search?
| Built-in web search | This MCP server | |
|---|---|---|
| Structured filters (price, form factor, condition) | guesswork | exact API filters |
| Uniform, comparable result list | no | yes |
| Repeatable / schedulable queries | no | yes |
| Fresh marketplace listings | often stale | live |
For one-off lookups the difference is small; for structured comparison and monitoring it is the point.
How it works
natural-language prompt
│
▼
LLM (OpenCode / Hermes / Claude) ── maps prompt → tool arguments
│
▼
MCP tools ── search, filters, details, references
│
▼
Marketplace providers ── OLX (JSON API), Hotline / Prom (SSR), Rozetka (Bright Data MCP)
│
▼
structured offers (title, price, location, url)
Tools
The server exposes these MCP tools (OLX / Hotline / Prom — registered out of the box):
search_offers(query, min_price, max_price, region_id, city_id, category_id, state, filters, limit, offset, include_promoted, sort)— one page of offers plus the total match count. Category-specific attributes go intofiltersas{"key": "value"}(choice) or{"key": [from, to]}(range).sort:newest(server-side),price_asc/price_desc(sorts the returned page). Promoted listings are excluded unlessinclude_promoted=True.get_offer_details(offer_id)— full listing: plain-text description, characteristics, photo URLs, seller, status, dates.list_filters(query, category_id)— discovers which category-specific filters exist for a query (keys + sample values). Call beforesearch_offerswhen an attribute is mentioned.category_idrestricts the sample to one category (broad queries otherwise mix categories).list_categories()— curated category id/name list (static snapshot; OLX deprecated the live category endpoint).list_regions()— OLX region (oblast) and major city ids.hotline_search_offers(query, min_price, max_price, limit, offset)— Hotline product models (new items from shops) with the minimum shop price.prom_search_offers(query, min_price, max_price, limit, offset)— Prom products (new items from shops).
Rozetka has no tool in this server — it is queried through the Bright Data MCP browser tools (see Optional: Bright Data MCP). The server's instructions tell the agent to do so automatically for price comparisons.
Example prompts
Find me an SSD 120–560 GB, 2.5", used, up to 1500 UAH on OLX
Cheap used monitors in Kyiv up to 2000 UAH, show the 5 cheapest
Details of offer 931436822
The agent maps these onto the tools: it calls list_filters to learn valid filter keys,
list_regions/list_categories for ids, then search_offers with the narrowed arguments.
Server-side instructions tell it to report only filters actually used and to format the
result as a list with prices and links.
Quick start
pip install marketua # or: uvx marketua (no install needed)
marketua # runs the MCP server over stdio (a client spawns it)
For development from this repo:
python3 -m venv .venv
.venv/bin/pip install -e ".[dev]"
.venv/bin/python -m pytest -q # tests
Connect an agent
Any MCP client works. OpenCode:
// opencode.json
{
"$schema": "https://opencode.ai/config.json",
"mcp": {
"marketua": {
"type": "local",
"command": ["uvx", "marketua"],
"enabled": true
},
// Rozetka (and any JS-rendered / Cloudflare-protected site) goes through
// the Bright Data MCP — free tier: 5 000 requests/month. Token from
// Bright Data → Settings → API Tokens (see .env.example in this repo).
"brightdata": {
"type": "remote",
"url": "https://mcp.brightdata.com/mcp?token={env:BRIGHT_DATA_API_KEY}&tools=scrape_as_markdown,scrape_as_html,search_engine&groups=browser",
"enabled": true
}
}
}
If the client can't find uvx: GUI apps (VS Code, Claude Desktop, Cursor) often don't inherit your shell's PATH, so a freshly installed uv (in ~/.local/bin) may be invisible to them. Find the path with which uvx and put it in command, or skip uvx entirely: pip install marketua and use marketua as the command.
Optional env vars. The server reads MARKETUA_DISABLE_PROVIDERS (comma-separated:
olx, hotline, prom) from its own environment to disable marketplaces. Most MCP clients
(Claude Desktop, Cursor — see below) forward a per-server env block from the config;
OpenCode does not forward env to spawned local MCP servers, so for OpenCode export the
variable in the shell before launching it.
Then ask e.g. "Find me a used monitor up to 2000 UAH". The model picks the right tool
itself: marketua_* for OLX / Hotline / Prom (free), brightdata_* for Rozetka and other
hard-to-reach sites.
Optional: Bright Data MCP (for Rozetka)
Rozetka is a Cloudflare-protected Angular SPA — it needs a real browser, so it goes through Bright Data's official MCP server. It is optional: without it the agent searches OLX / Hotline / Prom and says Rozetka was skipped; with it, price comparisons cover all four marketplaces.
- Free tier: 5 000 requests/month, no credit card. Token: Bright Data → Settings → API Tokens (not the proxy credential). The agent spends ~2–4 requests per search (navigate + snapshot + get_text).
- Data lives only in your chat — nothing is stored by us.
Claude Desktop (claude_desktop_config.json). The env block is optional —
include it only to disable marketplaces (comma-separated: olx, hotline, prom):
{
"mcpServers": {
"marketua": {
"command": "uvx",
"args": ["marketua"],
"env": {
"MARKETUA_DISABLE_PROVIDERS": "prom"
}
},
"brightdata": {
"url": "https://mcp.brightdata.com/mcp?token=YOUR_TOKEN&tools=scrape_as_markdown,scrape_as_html,search_engine&groups=browser"
}
}
}
Cursor (~/.cursor/mcp.json) — same optional env block for disabling marketplaces:
{
"mcpServers": {
"marketua": {
"command": "uvx",
"args": ["marketua"],
"env": {
"MARKETUA_DISABLE_PROVIDERS": "prom"
}
},
"brightdata": {
"url": "https://mcp.brightdata.com/mcp?token=YOUR_TOKEN&tools=scrape_as_markdown,scrape_as_html,search_engine&groups=browser"
}
}
}
For manual verification of every tool, use MCP Inspector:
npx @modelcontextprotocol/inspector --web uvx marketua
Platforms
| Marketplace | Status | Data source | Cost |
|---|---|---|---|
| OLX.ua | MVP (alpha) | public JSON API (api/v1/offers) |
free |
| Hotline.ua | MVP (alpha) | SSR search page | free |
| Prom.ua | MVP (alpha) | SSR search (embedded JSON-LD) | free |
| Rozetka.ua | via Bright Data MCP | remote browser (JS-rendered site) | free tier: 5 000 req/mo |
Architecture
One stable interface, many providers:
MCP layer (tools) → Provider interface → [OLXProvider, HotlineProvider, PromProvider]
(Rozetka — via the Bright Data MCP, no provider)
- MCP transport: stdio (simplest, the client spawns it as a child process). No hosted
server — shipped as a Python package (
pip install marketua/uvx marketua), so each user runs their own local instance. - Provider interface keeps marketplace specifics behind one abstraction, so adding a platform never touches the agent-facing layer.
- OLX category ids are a static snapshot (the live category endpoint is deprecated); the server detects stale ids at query time and warns instead of silently returning wrong results.
- Credentials (e.g. the Bright Data token) are configured by you in your client or shell, never committed.
License & disclaimer
MIT. This project is educational. It uses only public, unauthenticated data and respects
robots.txt and rate limits. Marketplace ToS may restrict automated access — use
responsibly, keep request rates low, and do not redistribute data at scale or bypass any
access controls. The optional Bright Data MCP is a third-party service with its own ToS and
free-tier limits.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。