TrendKia MCP Server

Version 1.2.1

An MCP server that exposes TrendKia — a bilingual Hindi/English news site — to AI assistants (Claude Desktop, claude.ai custom connectors, Claude Code, Cursor, etc.).

The site publishes a clean .md version of every article, so this server serves clean text — no HTML scraping.

Connect to the hosted server

A public, read-only instance is already running. Nothing to install:

https://mcp.trendkia.com

• claude.ai (Pro and above) — Settings → Connectors → Add custom connector, paste the URL.

• Claude Code — claude mcp add --transport http trendkia https://mcp.trendkia.com

• Claude Desktop / Cursor — add it as a remote HTTP MCP server in the client's config.

https://trendkia.com/mcp remains live for anything already pointed at it.

No API key or auth: it serves only content that is already public on the site.

Related MCP server: urltomarkdown-mcp

Tools

Languages

TrendKia publishes in Hindi at the site root, with an English edition under /en. There is no /hi prefix — a bare path is the Hindi article:

https://trendkia.com/market/<slug>-<id>        Hindi
https://trendkia.com/en/market/<slug>-<id>     English

The URL decides the language, and lang only overrides it.

• get_article(url) returns the article at that URL, in the language that URL is in. A URL you pass is never silently swapped for another edition.

• get_article(url, lang="en") returns the other edition of the same article — the server rewrites the path, so you never edit a URL yourself.

• search_articles and list_recent_articles default to Hindi, the site's own language, and take the same lang to switch.

Every result carries Language: and Alternate (…):, so the other edition is one field away in any single response — and the alternate is a clean canonical URL, never an echo of whatever query string you passed in.

The search index is shared, not per-locale — one full-text index over Hindi and English headlines, summaries and tags. So lang chooses which language the results come back in, not which index is consulted. That is why an English query matches Hindi articles.

Search vs. recent — the useful distinction

To find an article, use search_articles. It queries the site's full-text index and reaches every published article, however old. list_recent_articles reads the RSS feed and is limited to roughly the last 30 posts by design — it answers "what's new", nothing else.

list_sitemap_urls enumerates what exists; it is not a search tool. It defaults to kind="articles" because the sitemap opens with ~6,900 bribe-ledger state and district pages: before 1.3.0 the first 6,995 entries contained no article at all, so at the 1000-URL ceiling every possible response was ledger pages and the archive looked unreachable through it. Page through with offset.

Before 1.1.0 search_articles also read the feed, so a query for a common word returned only a handful of same-day hits. If you are running an older copy, upgrade.

Run it yourself

Local (Claude Desktop, stdio)

pip install -r requirements.txt
python trendkia_mcp.py

Add to claude_desktop_config.json:

{ "mcpServers": { "trendkia": { "command": "python", "args": ["/full/path/to/trendkia_mcp.py"] } } }

Hosted (HTTP transport)

MCP_TRANSPORT=http PORT=8000 python trendkia_mcp.py

Serves http://<host>:8000/mcp. Put it behind a TLS reverse proxy and point clients at the public URL. Dockerfile and render.yaml are included.

Reverse-proxy gotcha: FastMCP rejects a non-localhost Host header as DNS-rebinding protection, so a proxy forwarding Host: your-domain.com gets Invalid Host header until that host is allowed. Set it via the environment rather than patching the code.

Config

Requirements

mcp is pinned below 2.0 on purpose: mcp 2.x renames FastMCP to MCPServer and changes the decorator API, so an unpinned install fails at import. Moving to 2.x is a migration, not a version bump.