The news MCP.

Built for context windows, not front pages.
Live news, deduplicated into events — no key required.

01 — Try it

Run a real query. No account.

Pick a use case — see the call and the event it returns. Or open the full playground to search the index yourself.

news  subject="OpenAI"  limit=3
{
  "total": 139,
  "events": [
    {
      "event_id": "evt_9f5b87c838a3ff5cae0dcd52fe680857",
      "headline": "OpenAI Unveils Custom AI Chip",
      "abstract": "OpenAI has announced the development of its own AI chip, designed to reduce dependence on...",
      "entities": [{"name": "OpenAI", "salience": 0.9}, {"name": "Nvidia", "salience": 0.8}],
      "size": 6,                  // articles in the cluster
      "source_count": 6,          // distinct domains
      "newsrooms": 6,             // independent outlets
      "language_count": 5,
      "event_type": "product.product_launch",
      "sector": "semiconductors",
      "confidence": 0.9,
      "first_seen": "2026-08-25T23:03:26",
      "last_seen": "2026-08-26T23:08:54",
      "sources": [
        "https://www.dt.co.kr/article/12080229"
        // + 2 more source link(s)
      ]
    }
    // + 9 more event(s) in this response
  ]
}

02 — The tools

Ask it four things.

One call each. One more on Enterprise.

What's happening with Acme?

news(subject="Acme")

The name is quoted for you, and the window widens.

What's big right now?

news(sort="trending")

Recency-weighted, so an hour-old story can outrank yesterday's bigger one.

Tell me everything about this story.

get_story(event_id)

Every source link, not just the first three.

Is this actually true?

check_coverage(claim)

Independent newsrooms, counted apart from the domains that reprinted it.

monitorsENTERPRISE

Standing queries push new events to your endpoint.

03 — License and limits

Full commercial license. Any plan.

Ship what you build on it — in production, keyless included.

No key Free key
Lookback 3 days 7 days
Calls per hour 20 50
Stories per call 5 20

04 — NewsMCP or Enterprise

When to use which.

NewsMCP is for agent runs. Enterprise is for systems that depend on news.

AD-HOC AI RUNS
newsmcp.com

Keyless, clustered events, token-lean. Reach for it over MCP or its API inside Claude, Cursor, or an agent loop. The go-to for exploration and one-off runs.

PRODUCTION WORKFLOWS

When a business process runs on news: 2.5B+ articles indexed across 7+ years, 40+ filters, 30+ data fields, SLAs, and SOC 2 Type II.

05 — FAQ

Questions, answered.

Do I really not need a key?

No key to start — keyless runs at 20 calls an hour, looks back 3 days, and returns up to 5 stories per call. A free key takes that to 50 calls an hour, 7 days and 20 stories, and paid plans go higher. Asking for more than the keyless allowance isn't an error: the search runs narrowed and tells you what it changed. check_limits reports the ceilings your key actually has, and calling it costs nothing.

Can I use the results commercially?

Yes. Every plan includes a full commercial license, keyless included — ship what you build from day one.

MCP or REST API — which do I use?

Both exist, and they are not the same surface. REST is the raw index — one endpoint, every filter. The MCP server adds what REST has no endpoint for: check_coverage works out how many independent newsrooms carried a claim, and subject= quotes a name and widens the window so that searching for a company actually returns the company. If your agent speaks MCP, use MCP.

How do I add it to Claude or Cursor?

Run claude mcp add --transport http newsmcp https://mcp.newsmcp.com/mcp, or add {"url": "https://mcp.newsmcp.com/mcp"} to your client's MCP config and restart. The Setup for agents button copies a one-line instruction you can paste straight into an agent, which points it at our skill file so it configures itself.

How fresh is the data?

New stories land in the index continuously and are searchable within minutes of publication.

Where does the data come from?

One real-time news index, built and run entirely in house by NewsCatcher — no third-party search API sitting in the loop.

What languages and countries do you cover?

Worldwide sources across dozens of languages, keyless included — the same story is clustered across languages, so a single event routinely carries sources in six or seven. There is no language, country or source filter on the search itself: scope with sector, event_type, content_type, or a corroboration floor like min_newsrooms instead.

How are duplicates handled?

Automatically. The same story from many outlets is clustered into one event, so your agent reads a single result instead of thirty near-identical articles.

What's in a response?

Story events, not a flat article list. Each event carries an event_id, a generated headline, abstract and ranked entities; the labels event_type, content_type and sector with a confidence; three separate corroboration counts — size (articles), source_count (domains) and newsrooms (outlets that reported independently); first_seen/last_seen; and source links. verbosity sets how many links come back, and get_story returns all of them. Over MCP you get a readable markdown digest by default; pass response_format: json when you need to parse it. The REST endpoint has no such option — it always returns JSON.