Grounding chatbots with real-time news: stop stale and hallucinated current events

Ask an LLM what happened in the news this morning and you'll get a confident answer that's stale, vague, or invented. Its training stopped months ago — it has no idea what's breaking right now. The fix isn't a bigger model. It's grounding answers in a live news feed at query time.

👉 See the live data first — no signup, no key

Open it in your browser: api.newsagentdata.com/public/breaking →

You'll see the current breaking headlines as live JSON, each already carrying an urgency_score (0–10), political_lean, topic_tags, country_tags and an event cluster_id — the exact shape of data you'll hand your model, with no NLP pipeline of your own. Prefer the terminal?

curl https://api.newsagentdata.com/public/breaking

Why chatbots get current events wrong

A language model only knows what it was trained on. Anything after its cutoff — today's news — is missing or hallucinated with total confidence. For any assistant that touches current affairs, markets, politics or safety, that's a real risk.

Ground answers with retrieval

Retrieval-augmented generation (RAG): when a user asks something time-sensitive, fetch the relevant recent articles and inject them into the prompt as context, so the model answers from real, dated sources instead of memory.

curl -H "X-API-Key: YOUR_KEY" \
  "https://api.newsagentdata.com/v1/search?q=USER_TOPIC&days=2"

What the model actually gets

Every article comes pre-scored and classified — urgency_score (0–10), political_lean, topic_tags, country_tags, a timestamp, and an event cluster_id. So the model can rank what matters, cite the source and time, show balanced framing, and avoid repeating one event forty times. Coverage is deepest for English and Russian, and it reads ~3,000 public Telegram channels on top of RSS — sources web-only tools miss.

Keep the context fresh

Retrieve on-demand for a chatbot, or keep a vector store current with a webhook / SSE stream so new articles arrive within ~60 seconds. Either way the model never answers from stale memory.

Wire it up in a minute

The free tier is a real one — 100 requests/day of the full enriched schema, no credit card:

• 🔑 Free API key: https://newsagentdata.com/signup/?plan=free
• 🗺️ See coverage live (world map): https://newsagentdata.com/sources/map/
• 📚 Docs & full field list: https://newsagentdata.com/documentation/

The MCP server lets an agent query news as a tool, no glue code:

newsagentdata / newsagent-mcp

MCP server for the NewsAgent Data API — scored, classified Russian & English news intelligence for AI agents (Claude, Cursor, Cline).

NewsAgent Data — MCP server

Query scored, classified Russian & English news from any MCP client (Claude Desktop, Cursor, Cline, etc.). The server wraps the NewsAgent Data API — every article it returns is already scored for urgency (0–10), classified by political lean, topic, country and audience, and de-duplicated.

Tools

Tool	What it does
get_feed	Filtered feed — by country, topic, language, political lean, audience, min urgency, date range
search_news	Full-text keyword search across the archive
get_breaking	Recent high-urgency news (urgency ≥ 7 by default)
coverage_stats	Live totals — articles, sources, countries, languages (no key)
list_sources	Source catalog metadata (Standard tier+)

Setup

1. Get a free API key: https://newsagentdata.com/signup/?plan=free
2. Install deps: pip install -r requirements.txt
3. Add to your MCP client config. Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "newsagent": {
      "command": "python"
      "args": ["/absolute/path/to/newsagent_mcp.py"],
      "env": { "NEWSAGENT_API_KEY": 

…

View on GitHub

Building something and need higher volume or custom coverage? Message us on Telegram — @NewsAgentDatabot.

Honest note

Grounding sharply reduces but doesn't fully eliminate hallucination — keep the citations visible so users can verify. Russian and English are the deepest-enriched languages.

Originally published on NewsAgent Data.